{"id": "32a1ffc8bdc5bf86", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Reinforcement learning based process optimization and strategy development in conventional tunneling; DOI: 10.1016/j.autcon.2021.103701; license: cc-by."}, {"role": "assistant", "content": "As the present paper should be the first introduction of RL into tunneling we refrained from over-complicating the simulation and rather give an impetus for future studies that build upon this work. The below given improvements are a non-exhaustive list of ideas that we think are worth to be further explored. Ever improving processing power will alleviate problems related to computational cost over time. Nevertheless, all improvements must consider that each step in the framework will be done millions of times throughout the training and state of the art RL is by itself already computationally heavy. Improvements address either the agent or the environment: ▪ On the environment’s side, improvements could work towards making the geotechnical scenario more realistic by involving more ground types and other phenomena like groundwater and in-situ stress conditions. The excavation geometry could become more complex so that also other excavation shapes and sequences such as full-face excavation, or a further division of the top heading and bench excavation are possible. ▪ Closely related to the excavation geometry – and in our view one of the main points that should be improved – is the stability evaluation of the simulation. The reason why we used Eq. (3) after [29] as a substitute for more sophisticated means of sta­ bility assessment (e.g. tunnel cross sectional analyses such as analytical convergence confinement methods or 2D and 3D finite element analysis) is that this analytical solutions is computationally very efficient and does not prolongate the training process too much. For example, 100,000 episodes of the current framework with around 120 moves per episode would require ~1.2e+7 FEM based stability assessments. However, this reduction of the stability assessment is only us­ able for the ground conditions described in [29] and also ne­ glects other phenomena like the 3D stress state at the tunnel face. Improvements could be to do the stability assessment based on stand-up time concepts that also involve the rockmass quality [50] or ideally 3D finite element analysis as given above. ▪ Improvements for the agent on the one hand address the agent’s performance by modifying the agent’s architecture itself. Sys­ tematic hyperparameter tuning was not conducted yet and may Fig. 9. Histograms showing the performance of the agent 4 that was tested for 10,000 episodes in Fig. 7. Note the different scales of the x-axes in comparison to the histograms of Fig. 8. Fig. 10. Boxplot that shows how many times the agent 4 uses each action throughout 10,000 test episodes. Bold black lines in the boxes represent median values; the boxes confine the upper and lower quartiles and the whiskers show min.-max. Values; “al” in the x-labels refers to “advance length”. G.H. Erharter et al.\n[page 11] Automation in Construction 127 (2021) 103701 11 help to further improve the agent’s performance (e.g. grid−/ random search or even RL based hyperparameter optimization [51]). From a geotechnical point of view though, interesting improvements mainly concern an extension of the agent’s ca­ pabilities by introducing more possible actions. Whereas direct improvements to the given framework would be more advance lengths and types of tunnel support, other ideas are to involve exploration ahead of the face by simulated measurement while drilling [52] or geophysical exploration [53] to give the agent an idea what might be in front of the current excavation face. An idea in this regard is also to extend the agent to a multi- agent framework as it was used successfully before [15] where different agents have different tasks to fulfill. Future studies will work towards a more realistic environment and more complex agents in the TunnRL framework. Where TunnRL fits well in the line of the current development of automation, the greatest po­ tential lies in the possibility to develop new and not yet considered tunneling strategies for sequential tunnel excavation. Declaration of Competing Interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Appendix Appendix 1: Link to the GitHub repository where this paper’s code can be found: https://github.com/geograz/Tunnel-automation-with-Re inforcement-Learning-TunnRL- References [1] T. Marcher, G.H. Erharter, M."}], "source_title": "Reinforcement learning based process optimization and strategy development in conventional tunneling", "source_doi": "10.1016/j.autcon.2021.103701", "source_url": "https://www.sciencedirect.com/science/article/pii/S0926580521001527", "source_license": "cc-by", "zotero_item_key": "RXHUVV3X"}
{"id": "2b6282c18bad40b2", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Intelligent Classification of Surrounding Rock of Tunnel Based on 10 Machine Learning Algorithms; DOI: 10.3390/app12052656; license: cc-by-sa."}, {"role": "assistant", "content": "2022, 12, 2656. https://doi.org/10.3390/app12052656 https://www.mdpi.com/journal/applsci\n[page 2] Appl. Sci. 2022, 12, 2656 2 of 20 results. With the development of information technology, artificial intelligence, the Internet of Things, big data, and other technologies, the trend of tunnel construction automation and unmanned is becoming more and more obvious, and the intelligent classification of surrounding rock is the most important part. Thus, an automatic, rapid and accurate intelligent classification technology of surrounding rock is required for the construction of tunnel with high quality and high efficiency. Machine learning is an important method of artificial intelligence research [7–9] that has been applied in many types of tunnel analyses, such as deformation prediction [10–14], prediction of energy consumption of cutter head drives [15], rock burst prediction [16], reliability analysis [17,18], stability analysis [19], optimization of blasting parameters [20], support pattern selection [21], the prediction of blast-induced ground vibrations [22], tunneling risk prediction and assessment [23,24], diagnosing tunnel collapse sections [25], and TBM tunneling construction and management [26–30]. Machine learning is also an important method in the intelligent classification of surrounding rock. The physical and mechanical parameters of rock mass have been applied to the RMR value prediction using a neural network [31,32]. These parameters include the bulk density, compressive strength, ingress of water, rock quality designation (RQD), average distance between leak, and seismic velocity. The prediction of RMR has also been realized by using a neuro-fuzzy inference system based on the uniaxial compressive strength, RQD, joint or discontinuity spacing, joint condition, and groundwater condition [33]. In addition, the geophysical parameters, such as the seismic velocity and resistivity [34,35], have been used to classify the surrounding rocks [36–38]. The machine learning algorithms used in these studies include a variety of single basic algorithms, optimization algorithms, and integrated algorithms. These previous studies all show that the appropriate machine learning algorithms have excellent performance in different fields when there is an internal connection between input index and output index and the number of samples is sufficient. Especially for the highly nonlinear problems, the machine learning method often has better performance and higher computational efficiency than traditional statistical analysis methods. More importantly, the machine learning methods have the intelligent characteristics of automatic analysis and continuous learning, which provides effective help for this study. However, such methods generally require manual field testing of classification indices, such as rock strength and rock mass integrity, followed by manual input into the system. None of these parameters applied to the intelligent classification of surrounding rocks can achieve real-time automatic collection in the tunneling process. The emergence of measure-while-drilling (MWD) technology provides a good solution to this problem. The correlation between the drilling parameters and the surrounding rock quality parameters was studied and explored by scholars as early as the 1960s and 1970s [39–43]. Using field experiments with statistical analysis, the correlations between drilling parameters and surrounding rock quality parameters, such as the uniaxial com- pressive strength [44–46], shear strength [47], Schmidt rebound hardness [48], cutting performance (Kerf angle d and specific energy) [49,50], RQD [51], and zones of volcanic weathering and decomposition grades [52] have been studied. In recent years, drilling parameters have been used for surrounding rock classification based on the Q method [53] or RMR method [54]. Although these previous studies have confirmed the existence of a relationship be- tween drilling parameters and the quality of surrounding rock, the correlation based on current research of the drilling parameters and the surrounding rock quality parameters mostly refers to a certain lithology. Furthermore, the samples do not cover common rock lithology and are not universal."}], "source_title": "Intelligent Classification of Surrounding Rock of Tunnel Based on 10 Machine Learning Algorithms", "source_doi": "10.3390/app12052656", "source_url": "https://www.mdpi.com/2076-3417/12/5/2656/pdf?version=1646383280", "source_license": "cc-by-sa", "zotero_item_key": "UGAERXYU"}
{"id": "714ad63f9a7730de", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Wheel Loader Scooping Controller Using Deep Reinforcement Learning; DOI: 10.1109/access.2021.3056625; license: cc-by-sa."}, {"role": "assistant", "content": "This evaluation underscores the importance of the expert knowledge required for excavation. By contrast, most previous solutions to the automation of the scooping task (1) do not generalize to different machines and pile environments (2) rely on prior knowledge of an expert oper- ator and (3) require accurate models of the machine [10]. We presented a method that outperforms the above using the power of RL. Our learned controller acts based on measured sensors and implements the transitioning between phases with fixed predetermined criteria, which were developed solely through the training in the simulation. Future work should consider testing the proposed controller on a full-size skid-steer loader, as well as tackling the full problem of earthmoving including navigating to the pile, attempting a scooping cycle, navigating to another site and dumping. ACKNOWLEDGMENT The authors would like to thank RoboTiCan for lending them the robot used in the study presented in this work. REFERENCES [1] S. Dadhich, U. Bodin, and U. Andersson, ‘‘Key challenges in automation of earth-moving machines,’’ Autom. Construct., vol. 68, pp. 212–222, Aug. 2016. [2] S. Dadhich, ‘‘A survey in automation of earth-moving machines,’’ Lulea Univ. Technol., Luleå, Sweden, Res. Rep., 2015. [3] S. Singh, ‘‘State of the art in automation of earthmoving,’’ J. Aerosp. Eng., vol. 10, no. 4, pp. 179–188, Oct. 1997. [4] R. Filla, M. Obermayr, and B. Frank, ‘‘A study to compare trajectory gener- ation algorithms for automatic bucket filling in wheel loaders,’’ in Proc. 3rd Commercial Vehicle Technol. Symp. Commercial Vehicle Technol. (CVT), 2014, pp. 588–605. [5] T. Tomatsu, K. Nonaka, K. Sekiguchi, and K. Suzuki, ‘‘Model predictive trajectory tracking control for hydraulic excavator on digging operation,’’ in Proc. IEEE Conf. Control Appl. (CCA), Sep. 2015, pp. 1136–1141. [6] O. Luengo, S. Singh, and H. Cannon, ‘‘Modeling and identification of soil-tool interaction in automated excavation,’’ in Proc. IEEE/RSJ Int. Conf. Intell. Robots Systems. Innov. Theory, Pract. Appl., Oct. 1998, pp. 1900–1906. [7] K. Althoefer, C. P. Tan, Y. H. Zweiri, and L. D. Seneviratne, ‘‘Hybrid soil parameter measurement and estimation scheme for excavation automa- tion,’’ IEEE Trans. Instrum. Meas., vol. 58, no. 10, pp. 3633–3641, Oct. 2009. [8] D. Schmidt, M. Proetzsch, and K. Berns, ‘‘Simulation and control of an autonomous bucket excavator for landscaping tasks,’’ in Proc. IEEE Int. Conf. Robot. Autom., May 2010, pp. 5108–5113. [9] S. G. Olsen and G. M. Bone, ‘‘Modelling of robotic bulldozing operations for autonomous control,’’ in Proc. 24th Can. Conf. Elect. Comput. Eng. (CCECE), 2011, pp. 001188–001193. [10] S. Dadhich, F. Sandin, U. Bodin, U. Andersson, and T. Martinsson, ‘‘Field test of neural-network based automatic bucket-filling algorithm for wheel- loaders,’’ Autom. Construct., vol. 97, pp. 1–12, Jan. 2019. [11] I. Popov, N. Heess, T. Lillicrap, R. Hafner, G. Barth-Maron, M. Vecerik, T. Lampe, Y. Tassa, T. Erez, and M. Riedmiller, ‘‘Data-efficient deep rein- forcement learning for dexterous manipulation,’’ 2017, arXiv:1704.03073. [Online]. Available: http://arxiv.org/abs/1704.03073 [12] L. Tai, G. Paolo, and M. Liu, ‘‘Virtual-to-real deep reinforcement learning: Continuous control of mobile robots for mapless navigation,’’ in Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst. (IROS), Sep. 2017, pp. 31–36. [13] J. Hwangbo, I. Sa, R. Siegwart, and M. Hutter, ‘‘Control of a quadrotor with reinforcement learning,’’ IEEE Robot. Autom. Lett., vol. 2, no. 4, pp. 2096–2103, Oct. 2017. [14] S. Gu, E. Holly, T. Lillicrap, and S. Levine, ‘‘Deep reinforcement learning for robotic manipulation with asynchronous off-policy updates,’’ in Proc. IEEE Int. Conf. Robot. Autom. (ICRA), May 2017, pp. 3389–3396. [15] J. Tan, T. Zhang, E. Coumans, A. Iscen, Y. Bai, D. Hafner, S. Bohez, and V. Vanhoucke, ‘‘Sim-to-real: Learning agile locomotion for quadruped robots,’’ 2018, arXiv:1804.10332. [Online]. Available: http://arxiv. org/abs/1804.10332 [16] T. P. Lillicrap, J. J. Hunt, A. Pritzel, N. Heess, T. Erez, Y. Tassa, D. Silver, and D. Wierstra, ‘‘Continuous control with deep reinforce- ment learning,’’ 2015, arXiv:1509.02971. [Online]. Available: http://arxiv. org/abs/1509.02971 [17] O. Azulay. Package for Simulating Wheel Loader in Gazebo Engine for Reinforcement Learning Purposes. Accessed: Jan. 12, 2021. [Online]. Available: https://github.com/osheraz/komodo [18] P. J. A. Lever and F.-Y."}], "source_title": "Wheel Loader Scooping Controller Using Deep Reinforcement Learning", "source_doi": "10.1109/access.2021.3056625", "source_url": "https://ieeexplore.ieee.org/ielx7/6287639/9312710/09344588.pdf", "source_license": "cc-by-sa", "zotero_item_key": "IDZBIHT7"}
{"id": "64ba66f160885ff1", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Detection of Ionospheric Scintillation Based on XGBoost Model Improved by SMOTE-ENN Technique; DOI: 10.3390/rs13132577; license: cc-by-sa."}, {"role": "assistant", "content": "Overall, the accuracy of XGBoost remained higher than that of the decision tree and random forest, which presented similar accuracy at different tree depths. Meanwhile, the high accuracy of XGBoost started to stabilize when the tree depth reached 6, while there was a significant advantage in accuracy compared with the other methods. The higher accuracy of XGBoost with a lower tree depth means a simpler trained model and less running time, compared with the random forest, which relies on a larger tree depth and is more computationally expensive. Even though the decision tree model shows a high training efficiency, the accuracy was significantly lower than that of XGBoost, even when the tree depth reached 20 or higher, causing overfitting on the validation dataset and affecting the detection performance for novel data. The XGBoost algorithm proved to be appropriate, with high accuracy and low computational expense. Other, more vital parameters were set to default because there were no huge training data or large feature classes. Considering the running time and detection accuracy, the tree depth of XGBoost was set to 3 to reduce overfitting and improve the efficiency. The number of samples is also important for dealing with overfitting and underfitting problems. Based on the overall 349,559 points of the training set, 30% of the data were randomly reserved as an additional validation set, while 70% of the data were the training set, divided into 10, and the training size ranged from 0.1 to 1. As shown in Figure 7, the training accuracy across ten cross-validations and testing accuracy was compared at 10 different ratios for the training set. With an increase in training samples, the validation accuracy and testing accuracy gradually approached the same value and presented lower differences, which means that the model showed low bias and variance. Moreover, when the ratio of the training samples was set to 0.4 or higher, the accuracy variation decreased slightly and presented little fluctuation. To prevent overfitting or underfitting, the ratio was set to 0.4, meaning that approximately 100,000 points would be trained in following experiments. 4.3. Performance with Different Resampling Algorithms Aiming at mitigating the impact of data imbalance on the detection of strong scintil- lation, a combination of oversampling and undersampling algorithms was proposed to achieve resampling. As shown in Table 4, comparisons were made between the SMOTE- ENN algorithm and other several single oversampling or undersampling methods, namely, the random oversampling, random undersampling, SMOTE, and ENN algorithms, on the basis of the raw data used in the training process. All of the comparison experiments were\n[page 13] Remote Sens. 2021, 13, 2577 13 of 22 based on 100,000 points of data randomly selected from the overall training dataset, and the ratio of weak, moderate and strong scintillation events was 4.05:3.54:1, in accord with those in the overall training dataset. The training process was achieved with the XGBoost algorithm with a tree depth of 3. The raw data were trained directly on the XGBoost model, while the other methods required resampling on the raw data and then training on the XGBoost model. All the trained models were tested on the 38,841 points of the dataset mentioned in Table 1 with a ratio of weak, moderate and strong scintillation events equal to 3.30:2.56:1, including 11 segments of one-hour data that were not involved in the training phase. Accuracy (%) Figure 7. Training accuracy and testing accuracy of XGBoost versus the sample number used in the training set. In the overall training dataset mentioned in Table 1, 30% of the samples were randomly selected as the validation set, and the reserved samples were randomly selected as the training datasets according to different training sizes. Table 4. Performance comparison for XGBoost algorithm improved by different resampling techniques based on testing results."}], "source_title": "Detection of Ionospheric Scintillation Based on XGBoost Model Improved by SMOTE-ENN Technique", "source_doi": "10.3390/rs13132577", "source_url": "https://www.mdpi.com/2072-4292/13/13/2577/pdf?version=1625212202", "source_license": "cc-by-sa", "zotero_item_key": "4YEIWMK3"}
{"id": "afe2089e96ed2b5b", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Two-phase multi-model automatic brain tumour diagnosis system from magnetic resonance images using convolutional neural networks; DOI: 10.1186/s13640-018-0332-4; license: cc-by-sa."}, {"role": "assistant", "content": "Recently, deep learning paradigms have received con- siderable research attention. Deep learning models have multiple layers that can be trained using either supervised or unsupervised training approaches [9, 10]. In recent studies, deep learning models have achieved competi- tive results compared to the classical learning models in brain tumour diagnosis and tumour image classification [11, 12]. In addition, deep learning models like convolu- tional neural networks (CNNs) have improved the field of object detection and classification in different applications [13, 14]. As a deep learning paradigm, CNNs [15] have been used to extract high-level features from raw image data [16]. Although several studies have targeted the application of deep learning methods to brain tumour diagnosis, a complete system for automatic tumour detection and localization is not yet available in the literature. In addi- tion, integrating tumour detection and localization accu- racy within a single diagnosis system is still an open chal- lenge. Having brain tumour detection phase preceding the tumour localization in a single system leads to discarding the normal images from being processed in the localiza- tion phase. This opens doors to the real-time deployment of such automatic tumour diagnosis systems that save time and computing power from locating tumours in normal images. This research addresses the aforementioned challenges by developing a novel and complete computer-aided diag- nosis (CAD) system for tumour detection and localization from MRIs. In the tumour detection phase, the system combines a CNN, which is used for feature extraction due to its ability to learn features from raw data, with an error-correcting output codes support vector machine (ECOC-SVM), which is used for feature classification. The system is considered a two-phase multi-model arte- fact due to its detection and localization abilities using different CNN models. Three CNN models, namely, AlexNet, Visual Geome- try Group (VGG)-16, and VGG-19, have been deployed and evaluated for tumour detection. A five-layer fully developed region-based convolutional neural network (R- CNN) has been used in the second phase for tumour localization. The accuracy of the developed two-phase multi-model system has been evaluated using MRIs extracted from two different databases: the Reference Image Database to Evaluate Response (RIDER) Neuro MRI database [17] and the Brain Tumor Segmentation (BraTS) 2013 database. The key contributions of this study exist in several dimensions. First, a complete two-phase multi-model sys- tem for brain tumour detection and localization that combines a CNN and an ECOC-SVM is implemented and evaluated. Although CNNs and SVMs are not new approaches, the state-of-the-art in this field does not yet include systems similar to the one developed here, i.e., with these specific phases and CNN learning models. Sec- ond, integrating tumour detection and localization in one artefact is a good step towards real-time deployment of the research findings. Third, 349 MRIs from the RIDER Neuro MRI dataset are extracted and manually annotated, and this dataset will be available as a standard benchmark for other researchers who conduct research on tumour detection. Fourth, the conducted experimental works have proven the superiority of the developed system in terms of tumour detection and tumour localization. The remainder of this article is organized as follows. Section 2 is devoted to reporting related work on both the detection and localization of brain tumours. The description of the two-phase multi-model tumour diag- nosis system with respect to design and implementation is provided in Section 3. The experimental studies and the simulations for evaluating the whole diagnosis system in terms of tumour detection and localization accuracy are documented in Section 4. Finally, concluding remarks are outlined in Section 5. 2 Related work Brain tumour detection has received substantial research attention, and several detection methods have been pre- sented over the past two decades. In [18], a combina- tion of morphological filters, discrete wavelet transforms (DWTs), principal component analysis (PCA), and ker- nel support vector machines (KSVM) were applied to classify MRIs as normal and abnormal."}], "source_title": "Two-phase multi-model automatic brain tumour diagnosis system from magnetic resonance images using convolutional neural networks", "source_doi": "10.1186/s13640-018-0332-4", "source_url": "https://jivp-eurasipjournals.springeropen.com/track/pdf/10.1186/s13640-018-0332-4", "source_license": "cc-by-sa", "zotero_item_key": "2ZZSU54R"}
{"id": "488bc6c38dd23d2b", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Detection of Ionospheric Scintillation Based on XGBoost Model Improved by SMOTE-ENN Technique; DOI: 10.3390/rs13132577; license: cc-by-sa."}, {"role": "assistant", "content": "Conclusions Severe ionospheric scintillation is an adverse factor influencing the amplitude and carrier phase of a GNSS signal; its detection is a prerequisite in the design of an advanced re- ceiver with greater accuracy, reliability and efficiency. Nevertheless, the natural appearance of strong ionospheric scintillation occurs incidentally compared to that of weak/moderate scintillation. The imbalance may prove a challenge in achieving higher detection accuracy for strong scintillation events. As a strategy for detecting the severe ionospheric scintilla- tion events, the eXtreme Gradient Boosting (XGBoost) algorithm improved by the synthetic minority oversampling technique and edited nearest neighbor (SMOTE-ENN) resampling technique was developed as follows: (1) The machine learning method of XGBoost was proposed to improve the overall detection accuracy. According to 10 cross-validations, the accuracy was better than that of the decision tree and random forest. Meanwhile, XGBoost demonstrated sufficient validation accuracy when the tree depth was set to a small value, which not only significantly simplified the model complexity, but also effectively alleviated the overfitting problem. (2) Aiming at dealing with imbalance, different resampling techniques were compared based on the XGBoost detection model. SMOTE-ENN outperformed the other tech- niques on the whole. Moreover, similar improvements were observed for the decision tree and random forest detection models, after the SMOTE-ENN resampling tech- nique, while the improved XGBoost performed better than the other methods. (3) As for training datasets with different degrees of imbalance in classes 0/1/2, with ra- tios ranging from 1:1:1 to 7:7:1, different detection models and corresponding models improved by SMOTE-ENN were trained and then tested on the same novel dataset. The results showed overall enhancements for the improved detection methods com- pared to the corresponding raw methods, among which the improved XGBoost method showed the best performance. (4) The improved methods were tested with different degrees of imbalance in real data to evaluate the performance of the improved XGBoost. The results show distinct enhancements in overall accuracy, recall for class 2 and F1 scores for class 2, proving significant improvements in detecting severe scintillation events as well as reducing the problem of missing important events. Consequently, the performance of XGBoost improved by SMOTE-ENN was examined in comparative tests and under various conditions. The significance of these results lies in dealing with the problems of the natural and incidental appearance of strong scintillation\n[page 20] Remote Sens. 2021, 13, 2577 20 of 22 events, which may cause imbalance and especially affect the detection accuracy for strong scintillation events. This work would be of general interest for researchers in the fields of detecting interference in satellite signals (e.g., ionospheric scintillation, solar radio burst, and spoofing); the design of advanced receivers with greater accuracy, reliability and efficiency; and the atmospheric layer and space weather. Author Contributions: Conceptualization, M.L.; methodology, M.L.; software, X.Z.; validation, X.T. and G.T.; formal analysis, M.L.; writing—original draft preparation, M.L.; writing—review and editing, T.H.; supervision, X.C. All authors have read and agreed to the published version of the manuscript. Funding: This research was funded by the National Key Research and Development Plan of China, grant number 2018YFB0505103, and the National Natural Science Foundation of China, grant number 61873064. Data Availability Statement: Relevant data are available on request from the corresponding author. Please contact Xuefen Zhu (zhuxuefen@seu.edu.cn). Acknowledgments: The authors would like to thank three anonymous reviewers for their valuable and insightful comments. Conflicts of Interest: The authors declare no conflict of interest. References 1. Cai, X.; Burns, A.G.; Wang, W.; Qian, L.; Liu, J.; Solomon, S.C.; Eastes, R.W.; Daniell, R.E.; Martinis, C.R.; McClintock, W.E. Observation of Postsunset OI 135.6 nm radiance enhancement over South America by the GOLD mission. J. Geophys. Res. Space Phys. 2020, 126, 2020JA028108. 2. Karan, D.K.; Daniell, R.E.; England, S.L.; Martinis, C.R.; Eastes, R.W.; Burns, A.G.; McClintock, W.E. Early morning equatorial ionization anomaly from GOLD observations. J. Geophys. Res. Space Phys."}], "source_title": "Detection of Ionospheric Scintillation Based on XGBoost Model Improved by SMOTE-ENN Technique", "source_doi": "10.3390/rs13132577", "source_url": "https://www.mdpi.com/2072-4292/13/13/2577/pdf?version=1625212202", "source_license": "cc-by-sa", "zotero_item_key": "4YEIWMK3"}
{"id": "bb31d576cfbf8869", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Intelligent Classification of Surrounding Rock of Tunnel Based on 10 Machine Learning Algorithms; DOI: 10.3390/app12052656; license: cc-by-sa."}, {"role": "assistant", "content": "This is mainly because the use of an intelligent drill jumbo (which refers to a drill jumbo that can automatically collect and transmit drilling param- eters) for tunnel construction is low, which makes it more difficult to collect sufficient drilling parameter samples for the surrounding rock classification. Thus, the classification method with more objective, intelligent, and efficient evaluation requires further study.\n[page 3] Appl. Sci. 2022, 12, 2656 3 of 20 The purpose of this research is to introduce 10 machine learning algorithms to predict the quality of surrounding rock using MWD data (drilling parameters) obtained from five tunnels of the Zhengzhou–Wanzhou line of the high-speed railway project in China. Through comparative analysis, three machine learning models with better comprehensive performance among them were selected to establish the tunnel surrounding rock intelligent classification system by the drill and blast method. The results of this study lay a solid foundation for the dynamic design and intelligent construction of tunnels. 2. Materials and Methods 2.1. Proposed Methods and Procedures In this study, we firstly collected sample data, and then conducted a range of data processing, including sample data cleaning, sample imbalance treatment, sample feature analysis, and sample data splitting. Finally, we trained some intelligent classification models of the surrounding rock, and selected the better of them. The research flow chart about main procedures and proposed methods of this study is shown in Figure 1. Figure 1. The research flow chart about main procedures and proposed methods. 2.2. Sample Collection The sample of this study was obtained from some tunnels of the Zhengzhou–Wanzhou line of the high-speed railway project in China. This railway line runs from Zhengzhou East Railway Station to Wanzhou North Railway Station, with a total length of 818 km. It has 18 stations and a designed speed of 350 km/h, which connects Henan province, Hubei province, and Chongqing province. There are 32.5 tunnels in the Hubei province section, with a total length of 167.6 km, and the lithologies of stratum exposed by these tunnels are mainly dolomite, sandstone, limestone, shale, and mudstone. As shown in Figure 2, the\n[page 4] Appl. Sci. 2022, 12, 2656 4 of 20 sample collection was carried out in five tunnels in the Hubei province section. The New Austrian Tunneling Method was employed in these tunnels. Figure 2. Locations of Zhengzhou–Wanzhou high-speed railway line and the tunnels where the samples were collected. The sample of intelligent surrounding rock classification is composed of the drilling parameters and the surrounding rock grade of the tunnel face. In this study, the drilling parameters were collected by using the intelligent drill jumbo (Figure 3), which was made by China Railway Construction Heavy Industry Corpora- tion Limited. Figure 3. Intelligent drill jumbo (made by China Railway Construction Heavy Industry Corpora- tion Limited). The intelligent drill jumbo has the functions of automatic positioning, automatic drilling and automatic recording of log. It can complete the drilling operation of pipe roof, anchor bolt and blast hole. In the process, the integrated sensors of the intelligent drill jumbo are used to automatically collect the drilling parameters (penetration velocity, hammer pressure, rotation pressure, and feed pressure) during the drilling of the borehole in the tunnel face. The drilling parameters are described below: 1. Penetration velocity (Vp, m/min): the rate of penetration of the drill bit through the rock mass.\n[page 5] Appl. Sci. 2022, 12, 2656 5 of 20 2. Hammer pressure (Ph, bar): the measurement of the impact pressure of the bit against the rock mass. 3. Rotation pressure (Pr, bar): the pressure of the bit against the rock to maintain the required rotation. 4. Feed pressure (Pf, bar): the hydraulic pressure inside the cylinders required to keep the bit in contact with the bottom of the hole. All of these parameters are recorded at equal depth intervals of 20 mm by the help of displacement transducer. The drilling parameters data in this study were collected through the process of blasting hole drilling in the tunnel face (Figure 4a), and the bit used was a cemented carbide bit with a diamond content of 7%, spherical shape, and nine teeth (Figure 4b). (a) (b) Figure 4."}], "source_title": "Intelligent Classification of Surrounding Rock of Tunnel Based on 10 Machine Learning Algorithms", "source_doi": "10.3390/app12052656", "source_url": "https://www.mdpi.com/2076-3417/12/5/2656/pdf?version=1646383280", "source_license": "cc-by-sa", "zotero_item_key": "UGAERXYU"}
{"id": "64703313b8d12486", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Intelligent Classification of Surrounding Rock of Tunnel Based on 10 Machine Learning Algorithms; DOI: 10.3390/app12052656; license: cc-by-sa."}, {"role": "assistant", "content": "In particular, the recall of grade-V surrounding rock, which has great influence on the safety of the tunnel, was above 0.95. Thus, these three machine learning models have a degree of safety and high practi- cal value. 4. The classification accuracy of the surrounding rock intelligent classification system on-site is 86.7%, which indicates that these models have good generalization. This study found that the drilling parameters Vp, Ph, Pf, and Pr can be used to grade the surrounding rock directly by training the machine learning model. Compared with the traditional manual method, the proposed method is faster and has fewer indices, higher classification accuracy, and better stability. When carried on an intelligent jumbo, it can realize automatic recording and transmission of drilling parameters and intelligent classifi- cation of surrounding rock grade by using the intelligent surrounding rock classification system established in this study. This is what most previous studies have failed to do. However, there is still room to improve the classification accuracy of this system. In this study, three models have been used to determine the most unfavorable results to ensure the safety of site construction, and its essence is a management means. Therefore, the research on data preprocessing and algorithm optimization should be strengthened to further improve the reliability of the technology. In addition, the samples collected in this study only covered five lithologies and three surrounding rock grades. Thus, the sample types and numbers should be further expanded in the future to improve the application range of the technique. Author Contributions: Conceptualization, S.Z. and J.T.; methodology, S.Z. and J.T.; software, W.Y.; data curation, D.Y.; writing—original draft preparation, S.Z. and J.T.; writing—review and editing, M.W. and J.T.; visualization, D.Y.; funding acquisition, M.W. All authors have read and agreed to the published version of the manuscript. Funding: The authors are grateful for the support from the National Natural Science Foundation of China (51878567, 51878568) and China Railway Corporation Science and Technology Research and Development Program (K2018G014, K2020G035).\n[page 19] Appl. Sci. 2022, 12, 2656 19 of 20 Conflicts of Interest: The authors declare no conflict of interest. References 1. Zhang, X.; Wang, M.; Li, J.; Wang, Z.; Tong, J.; Liu, D. Safety factor analysis of a tunnel face with an unsupported span in cohesive-frictional soils. Comput. Geosci. 2020, 117, 103–221. [CrossRef] 2. Dla, D.; Yl, B.; Jc, C.; Hm, A. An analytical model for axial performance of rock bolts under constant confining pressure based on continuously yielding criterion. Tunn. Undergr. Space Technol. 2021, 113, 103955. 3. Li, D.; Cai, M.; Masoumi, H. A constitutive model for modified cable bolts exhibiting cone shaped failure mode. Int. J. Rock Mech. Min. 2021, 145, 104855. [CrossRef] 4. Barton, N.; Lien, R.; Lunde, J. Engineering classification of rock masses for the design of tunnel support. Rock Mech. 1974, 6, 189–236. [CrossRef] 5. Bieniawski, Z.T. The geomechanics classification in rock engineering applications. In Proceedings of the 4th ISRM Congress, Montreux, Switzerland, 2–8 September 1979. 6. Yunmei, L. Study of BQ formula in national standard of qunatitative classfication for basic quality of rock mass. Chin. J. Geotech. Eng. 1999, 4, 481–485. 7. Reich, Y. Machine learning techniques for civil engineering problems. Comput.-Aided Civ. Infrastruct. Eng. 1997, 12, 295–310. [CrossRef] 8. Elshaboury, N.; Abdelkader, E.M.; Al-Sakkaf, A.; Alfalah, G. Teaching-Learning-Based Optimization of Neural Networks for Water Supply Pipe Condition Prediction. Water 2021, 13, 3546. [CrossRef] 9. Elshaboury, N.; Mohammed Abdelkader, E.; Al-Sakkaf, A.; Alfalah, G. Predictive Analysis of Municipal Solid Waste Generation Using an Optimized Neural Network Model. Processes 2021, 9, 2045. [CrossRef] 10. Xue, X.; Xiao, M. Deformation evaluation on surrounding rocks of underground caverns based on PSO-LSSVM. Tunn. Undergr. Space Technol. 2017, 69, 171–181. [CrossRef] 11. Mahdevari, S.; Torabi, S.R. Prediction of tunnel convergence using artificial neural networks. Tunn. Undergr. Space Technol. 2012, 28, 218–228. [CrossRef] 12. Torabi-Kaveh, M.; Sarshari, B. Predicting Convergence Rate of Namaklan Twin Tunnels Using Machine Learning Methods. Arab. J. Sci. Eng. 2019, 45, 3761–3780. [CrossRef] 13. Qi, C.; Fourie, A.; Zhao, X."}], "source_title": "Intelligent Classification of Surrounding Rock of Tunnel Based on 10 Machine Learning Algorithms", "source_doi": "10.3390/app12052656", "source_url": "https://www.mdpi.com/2076-3417/12/5/2656/pdf?version=1646383280", "source_license": "cc-by-sa", "zotero_item_key": "UGAERXYU"}
{"id": "e7bf2eb49f7f3da9", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Reinforcement learning based process optimization and strategy development in conventional tunneling; DOI: 10.1016/j.autcon.2021.103701; license: cc-by."}, {"role": "assistant", "content": "The tunnel’s cross section has a total area of 91.31m2 and a height of 10 m with 58.56m2 and 32.75m2 being the areas of the top heading - and the bench and invert excavation respec­ tively (see Fig. 3). From these areas, equivalent diameters (D, i.e. the diameter of a circle with that area) of 8.41 m and 6.46 m can be computed for both parts of the excavation. There is a penalty if the distance between the top heading’s- and the bench’s tunnel face is too big. This distance (distmax) is set to be 50 m and the idea behind this is, that in many real projects – especially in soil conditions and long tunnels -, the top heading cannot be driven indefi­ nitely long ahead of the bench, because of safety reasons, necessities to Fig. 2. Simplified schematic plot of components of drill and blast tunneling (grey) as a reinforcement learning process - TunnRL. Fig. 3. Tunnel cross section of the given simulation. G.H. Erharter et al.\n[page 4] Automation in Construction 127 (2021) 103701 4 have a fast final ring closure, excavate cross cuts or general construction logistics. We defined two ground types (gt) for the simulation and the relevant ground properties are given in Table 1. Both are considered to be of homogeneous, isotropic and continuous nature and one can imagine them as a type of hard soil/soft rock (HSSR) material [30,31]. With the chosen parameters, gt1 represents “unfavorable” ground conditions and gt2 represents “favorable” ground conditions. The condition for a gt to be favorable/unfavorable is based on the ground properties in combi­ nation with the given tunnel geometry (see above) as evaluated by the chosen stability assessment criterion (see Section 3.3.2). The simulated tunnel is situated above the groundwater table and the permeability is set to 10−5 m/s for both gt. Choosing the given properties to define the gt, is closely connected to the way the stability assessment of the excavation is done (see Eq. (3) in Section 3.3.2). The permeability of 10−5 justifies the use of Eq. (3) for a stability assessment as according to [32] conditions are considered to be drained when the permeability is above 10−7 to 10−6 m/s. Nevertheless, if ground types with lower permeability/undrained conditions are to be used for the simulation, appropriate solutions for stability assessments must be chosen. Implementing more sophisticated stability assessments that take phenomena like ground water conditions into account is desirable but must be done with care as this heavily influences the overall performance of the tunneling simulation and the RL itself. The outlook and discussion of stability assessments in Section 7 goes into more detail on this topic. Using random walks with barriers [33], we created unique, 210 m long geological sections with a decimeter resolution (see Section 3.3.1 for why the sections are 210 and not 200 m long). We created a ground type-vector (gt-vector) of 2100 datapoints by scaling the random walk between 0 and 1, rounding to full numbers and using 0 as gt1 and 1 as gt2 (see Fig. 4 bottom row). To transform the gt-vector to a full geological section with one row for the top heading and one for the bench excavation, the vector is horizontally duplicated to an array of 2 × 2100 datapoints. + 1 is then added to the array so that the number 1 represents gt1, number 2 gt2 and number 0 represents the unexcavated part of the tunnel. The top row of Fig. 4 shows a visualization of such an array, where posth and posbi are at 165 and 125 m respectively. Given the complexity of simulating the process of conventional tunneling (see Section 2), we chose not to complicate the scenario by introducing more gt. However, increasing the number of gt can easily be done with the above described random walk based approach. For example, if it was necessary to simulate four gt then the values of the random walk must be split into four within the boundaries: gt1 < 0.25, 0.25 ≤gt2 < 0.5, 0.5 ≤gt3 < 0.75 and 0.75 ≥gt4. 3.2. Agent Translating the above described geotechnical scenario to RL, the geotechnician who observes the state of the construction and rockmass behavior and makes decisions based on this information, now becomes the “RL agent”. Due to the state-complexity of the problem, we chose a deep Q-network (DQN) as the RL agent."}], "source_title": "Reinforcement learning based process optimization and strategy development in conventional tunneling", "source_doi": "10.1016/j.autcon.2021.103701", "source_url": "https://www.sciencedirect.com/science/article/pii/S0926580521001527", "source_license": "cc-by", "zotero_item_key": "RXHUVV3X"}
{"id": "549a10457a09ad87", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Automated Excavator Based on Reinforcement Learning and Multibody System Dynamics; DOI: 10.1109/access.2020.3040246; license: cc-by-sa."}, {"role": "assistant", "content": "At the start of the training, during the loading stage, the excavator used a slightly higher angle of attack to the soil compared to the middle of the training, which resulted in a small inclination of the undercarriage when applying of the dipper arm force. Within the scope of this work, this behavior of the model is not dangerous, but in real-life applications, it can damage the machine. Therefore, in subsequent work, the reward function should penalize changes in the undercarriage rotation using data from an incli- nometer. Interesting behavior of the agent was also noticed prior to the addition of collision sensor data. Without the col- lision penalty, the agent tends to push with force on the hopper to achieve a higher reward. Complications with the reward for soil in the bucket are also worth mentioning. The agent had a tendency to abuse the reward for soil in the bucket and stop moving after filling the bucket. Therefore, the coefficient k with value of 0.2 was introduced to reduce the effect of the soil-based reward. An agent with a reward function, presented in (24), was able to grab soil and move it to the hopper, as shown in the plot of the average mass in the hopper in Fig.8. It can be clearly observed that the agent started moving ground at around 2 × 104 steps. The bucket is able to hold around 1200 kg of soil. As can be seen from Fig.9, on average the agent learnt to move 67% of the initial mass to the hopper. The maximum value of soil in the hopper was 1168 kg. Therefore, at maximum, the agent was able to deliver 97% of the initial mass in the bucket. In the Fig.8 and Fig.9 can be observed high standard deviation caused by the soil transfer FIGURE 8. Average and standard deviation of episode mass in the hopper. 214004 VOLUME 8, 2020\n[page 8] I. Kurinov et al.: Automated Excavator Based on RL and Multibody System Dynamics FIGURE 9. Average and standard deviation of episode mass in the bucket. FIGURE 10. Average and standard deviation of episode return. during simulation cycles, such as loading and unloading of the bucket. As shown in Fig.10, the average episode return increased over the course of the training, indicating that the agent was learning the policy correctly. Nevertheless, the high ampli- tude of the average reward shows that the agent did not reached the convergence. V. CONCLUSION This article introduced a method for development of autonomous machines based on reinforcement learning and a semi-recursive multibody method. The approach was tested on an excavator model with hydraulics and a deformable ground. The research concentrated on the learning agent used to operate the excavator model and move ground from a pre- defined point to the unloading point. The study used the PPO- CMA model-free algorithm. The algorithm and environment were connected via an API built based on the ZeroMQ library. The agent was able to learn a policy for working with the simulation models. The agent of the excavator was trained to load and unload the excavator with satisfactory accuracy. The agent was able to move 67-97% of the maximum mass of the bucket. Nevertheless, as was seen from the learning curves, the reward function still fluctuated. Thus, it can be concluded that increasing the number of episodes will likely result in enhanced excavator performance. The setup presented extends opportunities for development of automated machines using reinforcement learning. The multibody system dynamics allows the creation of machinery model and its environment with a little effort. This makes possible training of agents on the wide variety of machines. Therefore, the method allows highly customized simulation models to be created which can be used for training machine learning algorithms or neural networks. Hence, it will be possible to generate agents for existing machines effectively without endangering the machine or personnel. Furthermore, the approach provides a solid foundation for extension of RL learning to new types of machines, which can be easily reprogrammed by downloading agents to the machine model. REFERENCES [1] S. Dadhich, U. Bodin, F. Sandin, and U. Andersson, ‘‘From tele-remote operation to semi-automated wheel-loader,’’ Int. J. Electr. Electron. Eng. Telecommun., vol. 7, no. 4, pp. 178–182, 2018, doi: 10.18178/ijeetc.7.4. 178-182. [2] S. Dadhich, U. Bodin, and U. Andersson, ‘‘Key challenges in automation of Earth-moving machines,’’ Autom. Construct., vol. 68, pp. 212–222, Aug."}], "source_title": "Automated Excavator Based on Reinforcement Learning and Multibody System Dynamics", "source_doi": "10.1109/access.2020.3040246", "source_url": "https://ieeexplore.ieee.org/ielx7/6287639/6514899/09268069.pdf", "source_license": "cc-by-sa", "zotero_item_key": "7F2IQNFD"}
{"id": "0d5bf28bc1b5ea68", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: How to evaluate performance of prediction methods? Measures and their interpretation in variation effect analysis; DOI: 10.1186/1471-2164-13-s4-s2; license: cc-by."}, {"role": "assistant", "content": "VariBench is a database for variation-related benchmark datasets that can be used for developing, optimizing, com- paring and evaluating the performance of computational tools that predict the effects of variations (Nair and Vihi- nen, submitted). VariBench datasets provide multilevel mapping of the variation position to DNA, RNA and pro- tein as well as to protein structure entries in PDB [24] (when possible). Method developers are requested to sub- mit their datasets to VariBench to be distributed to the community. VariBench datasets have been collected from literature as well as with data mining approaches from diverse sources. Locus specific databases (LSDBs) are the most reliable source for disease-related data. Although lots of variation data are listed in LSDBs, it would be necessary to capture to databases all the cases from clinical and research laboratories [25,26]. An integral part of databases is the annotation of the entries. For variation information collection it would be extremely important to describe the cases in a systema- tic and unambiguous way. Variation Ontology (VariO, http://variationontology.org/ ) has been developed for systematic description and anno- tation of variation effects and consequences on DNA, RNA and/or protein including variation type, structure, function, interactions, properties and other features (Vihi- nen, in preparation). VariO annotated data would allow easy collection of novel dedicated benchmarks. Evaluation measures The outcome of binary (pathogenic/benign) style predic- tors are often presented in a 2x2 contingency table (Fig. 4). The number of correctly predicted pathogenic (non- functional) and neutral (functional) cases are indicated by TP (true positives) and TN (true negatives), and the num- ber of incorrectly predicted pathogenic and neutral cases are FN (false negatives) and FP (false positives), respectively. The goal of two-class prediction methods is to separate positive cases from negative ones. Because the predictions for the two classes usually overlap a cut off distinguishing the categories has to be optimized (Fig. 5). By moving the cut off different amounts of misclassified cases FN and FP appear. By using well behaved representative data and well trained classifier the misclassifications can be minimized. Figure 4 Contingency matrix and measures calculated based on it 2x2 contigency table for displaying the outcome of predictions. Based on the table it is possible to calculate row and column wise parameters, PPV and NVP, and sensitivity and specificity, respectively. These parameters are useful, but are not based on all the information in the table. Accuracy is a measure that is calculated based on all the four figures in the table. Figure 5 Separation of classes In most classification problems the two classes are overlapping. By moving the cut off position the amount of the overlap of the classes can be adjusted. FN and FP are misclassified cases. The prediction methods aim at optimizing the cut off and thereby adjusting the numbers in the contingency table. Vihinen BMC Genomics 2012, 13(Suppl 4):S2 http://www.biomedcentral.com/1471-2164/13/S4/S2 Page 5 of 10\n[page 6] Based on the four descriptors several further measures can be calculated (Fig. 4). Sensitivity, also called true posi- tive rate (TPR) or recall, and specificity (true negative rate, TNR) show the ratio of the pathogenic and neutral cases correctly identified by the programs. Positive predictive value (PPV) (also called precision) and negative predictive value (NPV) is the conditional probability that a patho- genic or neutral variant is predicted as pathogenic or neu- tral, respectively. The mathematical basis of these and other parameters have been discussed in detail [27]. A single parameter cannot capture all the information of the contingency matrix. Unless representative numbers of positive and negative cases are used, the values of NPV and PPV may be biased, even meaningless. The usual requirement is that the numbers be equal. Sometimes in literature the datasets are very skewed. Table 1 indicates the effect of the class imbalance. Results are shown in addition to equally distributed dataset also for analyses when there is ± 25 % or ±50 % difference in the total num- ber of negative and positive cases."}], "source_title": "How to evaluate performance of prediction methods? Measures and their interpretation in variation effect analysis", "source_doi": "10.1186/1471-2164-13-s4-s2", "source_url": "https://bmcgenomics.biomedcentral.com/counter/pdf/10.1186/1471-2164-13-S4-S2", "source_license": "cc-by", "zotero_item_key": "LKJJ9WPZ"}
{"id": "5cc15d3776937510", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Cutterhead and Cutting Tools Configurations in Coarse Grain Soils; DOI: 10.2174/1874836801711010182; license: cc-by."}, {"role": "assistant", "content": "Layout of the cutting tools on the spoke type cutterhead. (a) Initial cutterhead. (b) damaged disc cutters. \u0002\u0003\u0004\u0005\u0005\u0006\u0007\b \b \u0003 \b\u0005\u0006\u0007\b \u000e\u000f\u0004\u0010\u0011 \u0006\u0007\b \u0012\u0013\u0013 \u0003\u0006\u0007\b \u0014\u0013\u0015\u0016\u0006\u0010\u0011 \u0004\u000f\n[page 9] Cutterhead and Cutting Tools Configurations The Open Construction and Building Technology Journal, 2017, Volume 11 189 Fig. (8). Structure of the scraper. Fig. (9). Structure of the precut bits. \u0017\u0018\b \u0019\u001a\u001a \u0002\u0003 \u0004\u0005 \u0006\u0007 \u0003\u0003 \u0002\u0005 \u0017\u0018\b \u0019\u001a\u001a \b\u0005\u0005 \u0002\u0003\u0005 \u0017\u0018\b \u0019\u001a\u001a \u001b\u0002 \u0006\u0012\u0016\u0015\u0004\u0006 \u001b\u0003 \u0006\u0012\u0016\u0015\u0004\u0006\n[page 10] 190 The Open Construction and Building Technology Journal, 2017, Volume 11 Li et al. Fig. (10). Shell bit. Fig. (11). Center fish-tail bit. The rake and clearance angles of the scrapers with the attachment chip of 55 × 20 mm size are 15° and 10° respectively. The Type I precut bits are installed at the outer area of the cutterhead whereas the Type II precut bit mounted at the inner area of the cutterhead. The precut bits are 50 mm higher than the scrapers. The shell bits, which can be taken as precut bits, are mounted at the periphery of the cutterhead aiming at protecting the cutterhead and reducing frictions between the cutterhead and the soils. The fish-tail bit at the center of the cutterhead is 440 mm higher than the scrapers. (1) The Lot 14 of Beijing subway Line 4 is constituted of three twin running tunnels sectors and two underground \b\u0005\u0005 \u0017\u0018\b \u0019\u001a\u001a \u0017\u0018\b \u0019\u001a\u001a \u0003\u0003\u0005 \u0002\u0003\u0005\u0005\n[page 11] Cutterhead and Cutting Tools Configurations The Open Construction and Building Technology Journal, 2017, Volume 11 191 stations. The three twin running tunnels, mainly buried in sandy gravel and cobble ground with the maximum grain size of more than 200 mm, are 0.82 km, 1.47 km and 1.33 km in length respectively. The first running tunnel of 0.82 km long was finished with minor wears (see Fig. (12a)). Subsequent to repairing some cutter bits with surface welding, the second running tunnel, 1.47 km in length, was bored and after completing the second tunnel it was found that the scrapers were seriously damaged but the blades of the ripper teeth were basically remained (see Fig. (12b)). Before driving the third running tunnel of 1.33 km long, extensive repairs were performed for cutting tools, which later guaranteed the smooth completion of the third tunnel construction without changing tools. Fig. (12). Cutterhead used for Lot 14 of Beijing subway Line 4 and worn cutting tools. (2) The Lot 11 of Beijing subway Line 10 comprises two twin running tunnels 2.1 km and 0.7 km in length respectively, located in sandy gravel\\cobble ground. According to the geological investigation, the cobble with grain size more than 15 mm accounts for as much as 45% by volume and the SPT values are 60-65. After finishing the first running tunnel of 2.1 km long, it was observed that the wear amounts of the cutting tools were significantly less than the theoretical values. Subsequent to repairs of cutting tools with surface welding, the second running tunnel was also completed without interruption. (3) The same type shield machine was adopted in constructing a 0.82 km long tunnel in sandy gravel/cobble soils, of Lot 18 of Beijing subway Line 5. The smooth driving without replacing cutting tools was achieved, and the scrapers and the ripper teeth were still in good shape (see Fig. (13)). Fig. (13). Used scraper and ripper tooth in good shape. (a) Minor wears of cutting tools. (b) Worn scraper and ripper tooth.\n[page 12] 192 The Open Construction and Building Technology Journal, 2017, Volume 11 Li et al. (4) The same type shield machine was also utilized in digging a 1.02 km long running tunnel of Lot 9 of Shenyang metro Line 1. The running tunnel mainly passed through the sandy gravel/cobble soils with the maximum grain size of 100 mm. The tunnel was finished without changing cutting tools. Only a few wears of the cutterhead and most cutting tools were found except some seriously abraded tooth bits (see Fig. (14)). Fig. (14). Shield machine arrived and the used cutterhead. (5) The same type shield machine was also adopted in driving a 1.17 km long running tunnel of Lot 13 of Shenyang metro Line 1. The gravel/cobble with grain size more than 20 mm accounted for 20-30% of the total weight, and the maximum grain size discovered was 110 mm. The tunnel was completed without replacing cutting tools. After arrival of the shield machine, it was seen that most damages of the cutting tools were caused by collisions."}], "source_title": "Cutterhead and Cutting Tools Configurations in Coarse Grain Soils", "source_doi": "10.2174/1874836801711010182", "source_url": "https://openconstructionbuildingtechnologyjournal.com/VOLUME/11/PAGE/182/PDF/", "source_license": "cc-by", "zotero_item_key": "7KDIA2X2"}
{"id": "fdf03244feb100d5", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Spectral Angle Mapping and AI Methods Applied in Automatic Identification of Placer Deposit Magnetite Using Multispectral Camera Mounted on UAV; DOI: 10.3390/min12020268; license: cc-by-sa."}, {"role": "assistant", "content": "Thereafter, regions within each image known to be occupied by magnetite iron sand and non-magnetite beach sand are pinpointed so as to act as target and noise ground truths, respectively (Figure 7). Using the quantitative labelled segmentation maps as demonstrated in Figures 6 and 7, input datasets containing within them magnetite and non-magnetite six bands deep ground truths are ready to be employed in the training and testing of supervised ML and DL algorithms. The output variables are predicted or classified from the training database [10]. This means that algorithms use the ground truths and labelled segmentation maps to try and learn some patterns in the data during training, so as to implement these learnt patterns to the testing dataset [19]. This therefore provides results in relation to the learnt patterns (Figure 7). At 2 m flight elevation, the total dataset size was 998,400,000, derived from the product of image sizes (1300 × 1600), the number of multispectral bands (6), and the number of captured images (80). At 10 m flight elevation, the total dataset size was 399,360,000, derived from the product of image sizes (1300 × 1600), the number of multispectral bands (6), and the number of captured images (32). Lastly, at 20 m flight elevation, the total dataset size was 99,840,000, derived from the product of image sizes (1300 × 1600), the number of multispectral bands (6), and the number of captured images (8).\n[page 11] Minerals 2022, 12, 268 11 of 19 Figure 6. Extraction of 6-bands deep segmentation maps (binazized for easier visual representation) used to create a magnetite iron sand database. Noise refers to non-magnetite pixels. Segmentation maps at each UAV flight elevation are based on the 0.12, 0.13 and 0.17 cosine similarity thresholds at 2 m, 10 m and 20 m. This process is summarized in step 2⃝of Figure 7, as well as Appendix A. Figure 7. Proposed process overview with emphasis on how data extracted from spectral angle mapping cosine similarities is segmented, assigned ground truths, and used in training artificially intelligent algorithms in the classification of magnetite iron sand. Appendix A has extended explana- tions on this procedure.\n[page 12] Minerals 2022, 12, 268 12 of 19 These dataset sizes were, however, too large to run using the computational resources employed in this study (32 GB memory, Intel i7-8750H CPU @ 2.2 GHz processor, NVIDIA GeForce GTX 1060 graphics, Windows 10, 64 bits operating system). To counter this, Figure 8 demonstrates our AI training approach. At 2 m flight elevation, the input dataset size used to train each algorithm was 12,480,000, derived from the product of image sizes (1300 × 1600), and the number of multispectral bands (6). Since 80 of such datasets exist, 80 algorithms were trained, and the best performing model from the rest was chosen as the most viable. The same approach was applied at 10 m and 20 m flight elevations, the difference being the number of times (32 times and 8 times, respectively) training was performed, this number corresponding to the number of images captured. With this approach, all image datasets can still participate in the training of the various algorithms, the training times are quicker, and it requires less computational resources. Figure 8. Input dataset sizes used to train the machine learning and deep learning algorithms as well as selection of the best performing model at each UAV flight elevation. 3.3.1. Classification via Machine Learning Models The input dataset sizes for training and testing variables for each of the ML algorithms stood at 12,480,000 (from Figure 8) for each of the three flight elevation heights. Moreover, a 5-fold-cross-validation was used at all times in order to ensure that every observation (2,496,000 × 5) from the original dataset (12,480,000) has the chance of appearing in training and test sets. Therefore, this method generally results in a less biased model compared to other validation methods [10]. By using the MATLAB R2020b (manufactured by MathWorks, Natick, MA, USA) classification learner Machine Learning toolbox, multiple ML algorithms were evaluated. Classification performances were presented in terms of global accuracy, average per-class precision, and time taken to train the algorithm. These features are said to be amongst the most significant classification evaluation criteria [15]."}], "source_title": "Spectral Angle Mapping and AI Methods Applied in Automatic Identification of Placer Deposit Magnetite Using Multispectral Camera Mounted on UAV", "source_doi": "10.3390/min12020268", "source_url": "https://www.mdpi.com/2075-163X/12/2/268/pdf?version=1645347690", "source_license": "cc-by-sa", "zotero_item_key": "VPIE7856"}
{"id": "eb2a63ef8e761a6a", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Reinforcement learning based process optimization and strategy development in conventional tunneling; DOI: 10.1016/j.autcon.2021.103701; license: cc-by."}, {"role": "assistant", "content": "Winkler, Machine learning in tunnelling – capabilities and challenges, Geomechanik Tunnelbau 13 (2020) 191–198, https:// doi.org/10.1002/geot.202000001. [2] DAUB, BIM in Tunnelling: Digital Design, Building and Operation of Underground Structures. http://www.daub-ita.de/fileadmin/documents/daub/gtcrec4/gtcre c11v3_BIM_in_Tunnelling_05-2019.pdf, 2019 (accessed 9 October 2020). [3] L. Wagner, A. Kluckner, C.M. Monsberger, P. Wolf, K. Prall, W. Schubert, W. Lienhart, Direct and distributed strain measurements inside a Shotcrete lining: concept and realisation, Rock Mech. Rock. Eng. 53 (2020) 641–652, https://doi. org/10.1007/s00603-019-01923-4. [4] G.H. Erharter, T. Marcher, C. Reinhold, Application of artificial neural networks for underground construction – chances and challenges – insights from the BBT exploratory tunnel Ahrental Pfons, Geomechanik Tunnelbau 12 (2019) 472–477, https://doi.org/10.1002/geot.201900027. [5] G.H. Erharter, T. Marcher, C. Reinhold, Comparison of artificial neural networks for TBM data classification, in: Proceedings of the 14th International Congress on Rock Mechanics and Rock Engineering (ISRM 2019), Foz de Iguassu, Brazil, 2019. [6] G.H. Erharter, T. Marcher, C. Reinhold, Artificial neural network based online rockmass behavior classification of TBM data, in: Information Technology in Geo- Engineering, firstst ed. twentiethtwentieth, Springer, 2020, pp. 178–188. [7] R. Wu, Y. Fujita, K. Soga, Integrating domain knowledge with deep learning models: an interpretable AI system for automatic work progress identification of NATM tunnels, Tunn. Undergr. Space Technol. 105 (2020) 103558, https://doi. org/10.1016/j.tust.2020.103558. [8] S.K. Shreyas, A. Dey, Application of soft computing techniques in tunnelling and underground excavations: state of the art and future prospects, Innov. Infrastruct. Solut. 4 (2019), https://doi.org/10.1007/s41062-019-0234-z. [9] W. Zhang, R. Zhang, C. Wu, A.T.C. Goh, S. Lacasse, Z. Liu, H. Liu, State-of-the-art review of soft computing applications in underground excavations, Geosci. Front. 11 (2020) 1095–1106, https://doi.org/10.1016/j.gsf.2019.12.003. [10] S. Isam, Z. Wengang, Use of soft computing techniques for tunneling optimization of tunnel boring machines, Underground Space (2020), https://doi.org/10.1016/j. undsp.2019.12.001. [11] B.B. Sheil, S.K. Suryasentana, M.A. Mooney, H. Zhu, Machine learning to inform tunnelling operations: recent advances and future trends, Proc. Inst. Civ. Eng. (2020) 1–18, https://doi.org/10.1680/jsmic.20.00011. [12] Q. Zhang, Z. Liu, J. Tan, Prediction of geological conditions for a tunnel boring machine using big operational data, Autom. Constr. 100 (2019) 73–83, https://doi. org/10.1016/j.autcon.2018.12.022. [13] G.H. Erharter, T. Marcher, MSAC: towards data driven system behavior classification for TBM tunneling, Tunn. Undergr. Space Technol. 103 (2020) 103466, https://doi.org/10.1016/j.tust.2020.103466. [14] V. Mnih, K. Kavukcuoglu, D. Silver, A.A. Rusu, J. Veness, M.G. Bellemare, A. Graves, M. Riedmiller, A.K. Fidjeland, G. Ostrovski, S. Petersen, C. Beattie, A. Sadik, I. Antonoglou, H. King, D. Kumaran, D. Wierstra, S. Legg, D. Hassabis, Human-level control through deep reinforcement learning, Nature 518 (2015) 529–533, https://doi.org/10.1038/nature14236. [15] D. Silver, A. Huang, C.J. Maddison, A. Guez, L. Sifre, G. van den Driessche, J. Schrittwieser, I. Antonoglou, V. Panneershelvam, M. Lanctot, S. Dieleman, D. Grewe, J. Nham, N. Kalchbrenner, I. Sutskever, T. Lillicrap, M. Leach, K. Kavukcuoglu, T. Graepel, D. Hassabis, Mastering the game of go with deep Fig. 11. Exemplary log of one episode. The x-axis shows the number of blasts/moves the agent needed to achieve breakthrough (i.e. 105): Top row: Time-distance diagram showing the logged position of top heading excavation (solid line), bench excavation (dashed line) and the difference in between (grey area). Middle row: Record of the actions (see Table 3 for the corresponding action codes). Bottom row: cumulative reward throughout the episode. G.H. Erharter et al.\n[page 12] Automation in Construction 127 (2021) 103701 12 neural networks and tree search, Nature 529 (2016) 484–489, https://doi.org/ 10.1038/nature16961. [16] J. Shahrabi, M.A. Adibi, M. Mahootchi, A reinforcement learning approach to parameter estimation in dynamic job shop scheduling, Comput. Ind. Eng. 110 (2017) 75–82, https://doi.org/10.1016/j.cie.2017.05.026. [17] M."}], "source_title": "Reinforcement learning based process optimization and strategy development in conventional tunneling", "source_doi": "10.1016/j.autcon.2021.103701", "source_url": "https://www.sciencedirect.com/science/article/pii/S0926580521001527", "source_license": "cc-by", "zotero_item_key": "N5Q33JID"}
{"id": "d197930ea694efd0", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Spectral Angle Mapping and AI Methods Applied in Automatic Identification of Placer Deposit Magnetite Using Multispectral Camera Mounted on UAV; DOI: 10.3390/min12020268; license: cc-by-sa."}, {"role": "assistant", "content": "This suggests that the models are all highly effective at separating magnetite iron sand spectra from non-magnetite beach sand spectra from within the multispectral image scenes. Figure 10. Deep learning one dimensional convolution neural network confusion matrices at 2 m, 10 m and 20 m UAV drone elevation.\n[page 16] Minerals 2022, 12, 268 16 of 19 4. Discussion In a study by Gholami et al. [28], they applied independent component analysis (ICA) in remote sensing data processing based on satellite data to map an area and attained good results. However, differences between their method and ours are as follows; with the ICA method, they reduced the dimensionality of spectral bands from six bands to two bands. However, researchers mention that higher dimensionalities are relatively superior in deducing materials as they retain the original data structure [5–9]. Hence, our study applied SAM, which retains the full six bands data structure. In addition, our methodology employs a UAV drone, which allows for collection of variables, as well as higher spatial resolution data. For example, at 2 m flight elevation, our spatial resolution is 0.4 cm/pixel, as opposed to their fixed satellite resolution of 30 m/pixel. Considering our 30 m × 90 m study area, their spatial resolution is rather inadequate. Lastly, we employ supervised DL 1D CNN classification algorithms which are known to be state of the art in imagery data classification problems. As pointed out by Sharma and Sharma [10], supervised learning algorithms are superior to unsupervised ML methods such as ICA, especially in cases where the output variables (magnetite or non-magnetite) are known, as is the case in this study. Based on this study’s findings, it was realised that the SAM cosine similarity thresholds require slight adjustments relevant for each UAV flight elevation. With increase in elevation from 2 m, 10 m to 20 m, these thresholds vary from 0.12, 0.13 to 0.17, respectively. Though these SAM cosine similarities vary, the ability of the SAM in differentiating the targeted magnetite iron sand from non-magnetite beach sand remains relatively unchanged from visual analysis. This is demonstrated by the masks at each flight elevation which seem to flawlessly overlap the sought after magnetite iron sands (Figure 5), as well as the high ground truths-supported AI classification capabilities. With respect to the UAV drone post flight logs and the typical size of surface mines, it would be in the best interest of engineers to perform mineral identification practices at higher elevations such as 20 m. The reason being, the UAV is not as challenged in terms of power consumption, elapsed capturing time, number of images to capture, and area coverage, as at lower elevations (Table 1). Figure 11 is a summarized comparison between the ML and DL CNN model classi- fication capabilities as well as the length of time required to train each model. Based on global accuracies, the CNN models perform better than the ML models across all three flight elevations. With ML, the attained global accuracies were 85.7%, 78.6% and 85.7% at 2 m, 10 m and 20 m flight elevations, whilst with the CNN, attained global accuracies were 99.9%, 99.9%, 99.7%, respectively. A look into the attained per-class precisions shows that with the ML models, 84.5%, 83.4%, and 89. A total of 9% classification capabilities were attained at 2 m, 10 m and 20 m flight elevations. These model classification capabilities are different from the 99.8%, 98.7% and 99.4% attained via the CNN at 2 m, 10 m and 20 m, respectively. Therefore, it can be said that based on the global accuracies as well as the per-class precisions (Figure 11a), model capabilities show that the 1D CNNs employed in this study are far superior in classifying the magnetite iron sands, and non-magnetite beach sands, to the ML models. From Figure 11b, one is able to appreciate the differences in times required to train the AI models. With ML models, training times are shorter (5.8, 0.9 and 1.0 s at 2 m, 10 m and 20 m, respectively) compared to those of the CNN models (78, 45 and 68 s at 2 m, 10 m and 20 m, respectively). Though the ML model training times seem relatively more efficient, the time taken to train a model is not what governs its industrial applicability. Training duration merely communicates the time it took to turn an algorithm into a model via training.\n[page 17] Minerals 2022, 12, 268 17 of 19 (a) (b) Figure 11."}], "source_title": "Spectral Angle Mapping and AI Methods Applied in Automatic Identification of Placer Deposit Magnetite Using Multispectral Camera Mounted on UAV", "source_doi": "10.3390/min12020268", "source_url": "https://www.mdpi.com/2075-163X/12/2/268/pdf?version=1645347690", "source_license": "cc-by-sa", "zotero_item_key": "VPIE7856"}
{"id": "fe2e65c4e96bbdab", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: How to evaluate performance of prediction methods? Measures and their interpretation in variation effect analysis; DOI: 10.1186/1471-2164-13-s4-s2; license: cc-by."}, {"role": "assistant", "content": "Vihinen BMC Genomics 2012, 13(Suppl 4):S2 http://www.biomedcentral.com/1471-2164/13/S4/S2 Page 3 of 10\n[page 4] data should however never be used for training and test- ing as that would only indicate the capacity of the method to memorize examples, not its generalization potential – how well it performs on instances outside the training set. High quality benchmark datasets require meticulous data collection often from diverse sources and careful checking of the correctness of the data. Numerous measures have been developed to describe predictor performance, but no single measure captures all aspects of predictor performance. The measures mainly used, and how to interpret them will be discussed. Typi- cally prediction methods are used as classifiers to define whether a case has the investigated feature or not. Results of this kind of binary predictor can be presented in a 2x2 confusion table also called contingency table or matrix. This, at first glance may appear simple to interpret, but the contrary is the case, as various composite aspects have to be jointly taken into account. Benchmark criteria Benchmark can be defined as a standard or reference for evaluation, in this case prediction method performance. Benchmarks are widely used in computer science and technology. For example computer processor performance is tested with standardized benchmark methods. In bioin- formatics there are benchmarks e.g. for multiple sequence alignment methods already 1990’s [9]. Novel MSA con- struction methods are routinely tested with alignment benchmarks such as BAliBASE [10] HOMSTRAD [11], OxBench suite [12], PREFAB [13], and SABmark [14] . Other bioinformatic benchmarks include protein 3D structure prediction [15-17], protein structure and func- tion prediction [18], protein-protein docking [19] and gene expression analysis [20,21] benchmarks etc. Benchmark usage varies between different communities. For variation effect predictions, benchmarks have not been available and thus authors have used different datasets. The situation has changed only recently with the release of VariBench (http://bioinf.uta.fi/VariBench/) (Nair and Vihinen, submitted). To be useful a benchmark should fulfill certain criteria. These criteria vary somewhat between the domains, but there are also some common features (Fig. 3). The criteria laid by Gray originally for database systems and transac- tion processing systems are still valid [22]. Criteria for MSA [23] and variation data (Nair and Vihinen, sub- mitted) benchmarks have been defined. These include relevance, which means that the data have to capture the characteristics of the problem domain. Portability allows testing of different systems. Scaleability of the benchmark allows testing systems of different sizes, and simplicity means that the benchmark has to be understandable and thereby credible. Accessibility means that the benchmark has to be publicly available, solvability to set the level of the task on suitable level (not too difficult, not hoo hard), independence to guarantee that the benchmark has not Figure 3 Benchmark criteria Criteria for benchmarks in three different studies. VariBench is the database specifically designed for variation benchmark datasets. Vihinen BMC Genomics 2012, 13(Suppl 4):S2 http://www.biomedcentral.com/1471-2164/13/S4/S2 Page 4 of 10\n[page 5] been developed with tools to be tested, and evolution to keep the benchmark up-to-date during time. When considering the variation benchmarks, datasets should be large enough to cover variations related to a cer- tain feature or mechanism. For example in the case of mis- sense variations this means very large numbers of instances as there are altogether 150 single nucleotide changes which cause amino acid substitution. To have sta- tistical power several cases are needed. The required num- bers of cases increase exponentially as features are combined. Datasets have to be non-redundant and devoid of similar or greatly overlapping entries. This criterion relates to independence requirement of [23]. Datasets have to contain both positive (showing the investigated feature) and negative (not having effect) cases so that the capability of methods to distinguish effects can be tested. This may cause problems in data collection as some phenomena are very rare and only a few known cases may exist."}], "source_title": "How to evaluate performance of prediction methods? Measures and their interpretation in variation effect analysis", "source_doi": "10.1186/1471-2164-13-s4-s2", "source_url": "https://bmcgenomics.biomedcentral.com/counter/pdf/10.1186/1471-2164-13-S4-S2", "source_license": "cc-by", "zotero_item_key": "LKJJ9WPZ"}
{"id": "19e52aeb5ff2c778", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Detection of Ionospheric Scintillation Based on XGBoost Model Improved by SMOTE-ENN Technique; DOI: 10.3390/rs13132577; license: cc-by-sa."}, {"role": "assistant", "content": "Meanwhile, Figure 4 shows that the recall value for the majority class 0 remains above 93%, while that for class 1 keeps increasing and reaches over 98%. However, there is a distinct decline for class 2 from 87% to 74%, with an approximate 13% decline, which means that an increasing number of strong scintillation events go undetected. This indicates the great importance of balancing the minority class data in improving the accuracy of strong scintillation detection. ( ) 1f X (1) y (0) y 1( ) Tree X (1) y 2( ) Tree X ( ) 2f X (2) y ( ) kf X ( ) k y ( 1) k y − ( ) Treek X ( ) if X  Training dataset X   Node splitting Obj(1) Node splitting Obj(2) Node splitting Obj(k-1) Iterative boosting Figure 3. The schematic diagram of XGBoost algorithm. Table 2. The overall accuracy values for different sample proportions for classes 0/1/2. Sample Proportion (Class 0/1/2) 1:1:1 2:2:1 3:3:1 4:4:1 5:5:1 6:6:1 7:7:1 Overall Accuracy (%) 90.18 91.47 93.32 93.13 93.31 92.65 92.61 1:1:1 2:2:1 3:3:1 4:4:1 5:5:1 6:6:1 7:7:1 Sample proportion for class 0/1/2 70 75 80 85 90 95 100 class 0 class 1 class 2 13% decline Figure 4. Influence of imbalanced datasets on performance of XGBoost model. These seven training datasets with 100,000 samples were selected randomly in the overall training dataset listed in Table 1, according to different sample ratios for classes 0/1/2. The testing dataset for each detection model trained by the above training dataset is the same as that mentioned in Table 1. The purpose of resampling in this study was to balance the ionospheric scintillation events of different intensities, enhancing the detection accuracy for strong scintillation events. The normally used undersampling methods include random undersampling of the majority class, edited nearest neighbor (ENN) [43], Tomeklink [44], condensed nearest\n[page 10] Remote Sens. 2021, 13, 2577 10 of 22 neighbor (CNN) [45], and neighborhood cleaning (NCL) [46], while the typical oversam- pling methods include random oversampling of the minority class, the synthetic minority oversampling technique (SMOTE) [47], and Borderline-SMOTE (BSM) [48]. However, the former methods miss sample information through reducing the majority samples, and the latter methods add minority samples, causing a constant increase in useless information and an overfitting problem. The combinational algorithm SMOTE-ENN preserves the features of majority samples and increases the characteristics of minority samples, resulting in excellent classification performance for imbalanced datasets [49]. The basic idea of the SMOTE method is to carry out linear interpolation between neighboring minority class samples to synthesize new minority class samples, solving the problem of significant data overlap compared with random oversampling [47]. Further details can be described as follows: for each minority class sample xm ∈X (m = 1, . . . , M), find the k nearest neighboring samples of the same class K, and then, choose n samples in K according to the sampling rate n, mark them as ym,i ∈Ym (i = 1, . . . , n, m = 1, . . . , M), and finally achieve random linear interpolation on the lines between xm and ym,1, . . . , ym,n, respectively. The new built sample can be described as (10). Xnew,m = xm + r ∗(Ym −xm), m = 1, . . . , M (10) where r is the random coefficient ranging from 0 to 1, and Xnew,m is the vector including n new samples built by xm. In total, there are M ∗n new samples. Focusing on majority class samples, the ENN algorithm deletes the sample if there are two or more in the nearest three neighboring samples different from it [43]. However, the majority samples are near each other, which causes limited sample removal. The SMOTE-ENN method achieves oversampling on the minority class samples, firstly using SMOTE, and then finishes undersampling on the majority class samples to coalesce their advantages. Thus, the combination of SMOTE and ENN is utilized to balance data with few strong ionospheric scintillation events; its performance was evaluated and compared with that of other methods, such as SMOTE and ENN. Figure 5 shows the process of resampling, training and predicting, using the XGBoost algorithm improved by the SMOTE-ENN resampling technique."}], "source_title": "Detection of Ionospheric Scintillation Based on XGBoost Model Improved by SMOTE-ENN Technique", "source_doi": "10.3390/rs13132577", "source_url": "https://www.mdpi.com/2072-4292/13/13/2577/pdf?version=1625212202", "source_license": "cc-by-sa", "zotero_item_key": "4YEIWMK3"}
{"id": "b649337eed839885", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Cutterhead and Cutting Tools Configurations in Coarse Grain Soils; DOI: 10.2174/1874836801711010182; license: cc-by."}, {"role": "assistant", "content": "Nishitake, \"Earth pressure balanced shield machine to cope with boulders\", In: Rapid Excavation and Tunneling Conference Proceedings, 1987, pp. 552-572. [11] P.B. Dowden, and D.T. Cass, \"Shielded TBM’s. Matching the machine to the job\", In: Rapid Excavation and Tunneling Conference Proceedings, 1991, pp. 787-805. [12] W. Burger, \"Design principles for soft ground cutterheads\", In: Rapid Excavation and Tunneling Conference Proceedings, 2007, pp. 784-792. [13] M. Herrenknecht, and K. Bӓpler, \"The latest technology in mechanized tunneling-the design of the World’s largest EPB and slurry shield TBMs\", In: Underground Space-The 4 th Dimension of Metropolises, Proceedings of the World Tunnel Congress 2007 and 33 rd ITA/AITES Annual General Assembly, 2007, pp. 1169-1173. [14] W. Burger, and G. Wehrmeyer, \"The making of the Mixshield-Part 1\", In: Tunnels & Tunnelling International, 2008a, pp. 35-40. [15] W. Burger, and G. Wehrmeyer, \"Lifting the lid on Mixshield performance-Part 2\", In: Tunnels & Tunnelling International, 2008b, pp. 31-36. [16] C. Metzger, G. Colzani, and G. Irwin, \"Selection, design, and procurement of North America’s largest mixshield TBM for Portland, Oregon’s East side CSO tunnel\", In: Papid Excavation and Tunneling Conference Proceedings, 2009, pp. 1022-1036. [17] U. Maidl, and M. Comulada, \"Prediction of EPB shield performance in soils\", In: Rapid Excavation and Tunneling Conference Proceedings, 2011, pp. 1083-1091. [18] N. Bilgin, H. Copur, and C. Balci, \"Effect of replacing disc cutters with chisel tools on performance of a TBM in difficult ground conditions\", Tunn. Undergr. Space Technol., vol. 27, pp. 41-51, 2012. [http://dx.doi.org/10.1016/j.tust.2011.06.006] [19] D. Willis, \"Reducing the cutterhead wear at EPB tunnel boring machines\", Tunnel, vol. 2012, no. 7, pp. 17-25, 2012. [Online] Available from: http://www.tunnel-online.info/en/artikel/tunnel_Reducing_the_Cutterhead_Wear_at_EPB_Tunnel_Boring_Machines_1552388.html [Accessed: 1 st Aug. 2016]. [20] M. Havekost, C. Overby, J. Kabat, and P. Nelson, \"Use of a digger shield to successfully complete tunnel after ground conditions proved too adverse for a TBM\", In: Rapid Excavation and Tunneling Conference Proceedings, 2013, pp. 264-275. [21] F. Köppl, and K. Thuro, \"Cutting tool wear prognosis and management of wear-related risks for Mix-ShieldTBM in soft ground\", In: Proceedings of the 18 th International Conference on Soil Mechanics and Geotechnical Engineering, 2013, pp. 1739-1742. [22] R. Duhme, and T. Edelmann, \"Monitoring and Intervention-Free Inspection of TBM Excavation Tools\", Underground Singnapore 2014, 25 th–26 th September, 2014. [23] J. Roby, and D. Willis, \"Achieving fast EPB advance in mixed ground: A study of contributing factors\", In: North American Tunneling Proceedings, 2014, pp. 182-194. [24] G.P. Le, \"Selection of shield machines for Beijing Metro construction\", Mod. Tunneling Technol., vol. 40, no. 3, pp. 14-30, 2003. [In Chinese]. [http://dx.doi.org/10.3969/j.issn.1009-6582.2003.03.004] [25] L. Wang, \"Application of shield tunneling method in Beijing metro line 5\", Urban Rapid Rail Transit, vol. 17, no. 5, pp. 33-37, 2005. (In Chinese). [26] G.F. Liu, and T.T. Song, \"Analysis and study on wearing of cutting tools of shield machines used in Chengdu metro construction\", Tunnel Constr., vol. 27, no. 6, pp. 89-93, 2007. [In Chinese]. [http://dx.doi.org/10.3969/j.issn.1672-741X.2007.06.025] [27] G.L. Huang, \"Study on shield tunneling in Shenyang water-rich sandy gravel ground\", Railway Constr., no. 4, pp. 77-78, 2008. (In Chinese). [28] M.F. Zhang, D.J. Yuan, Q.F. Huang, C.B. Huang, and F. Zhang, \"Analysis of dynamic abrasion of shield cutters in sandy cobble stratum\", Chin. J. Rock Mech. Eng., vol. 27, no. 2, pp. 397-402, 2008. [In Chinese]. [http://dx.doi.org/10.3321/j.issn:1000-6915.2008.02.024] [29] G.Y. Wang, \"Shield equipment disposition for Chengdu metro\", Tunnel Constr., vol. 28, no. 4, pp. 501-504, 2008. [In Chinese].\n[page 19] Cutterhead and Cutting Tools Configurations The Open Construction and Building Technology Journal, 2017, Volume 11 199 [30] Q.Z. Wang, and S.K. Zhang, \"The analysis of the arrangement of cutters and the structure of the cutter carrier of the shield for projects of the Shenyang metro in the medium-coarse sand stratum\", Traffic Eng. Techno. National Defense, vol. 9, no. 4, pp. 41-43, 2011. [In Chinese]. [31] Q.S."}], "source_title": "Cutterhead and Cutting Tools Configurations in Coarse Grain Soils", "source_doi": "10.2174/1874836801711010182", "source_url": "https://openconstructionbuildingtechnologyjournal.com/VOLUME/11/PAGE/182/PDF/", "source_license": "cc-by", "zotero_item_key": "7KDIA2X2"}
{"id": "51576c24dcde8655", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: How to evaluate performance of prediction methods? Measures and their interpretation in variation effect analysis; DOI: 10.1186/1471-2164-13-s4-s2; license: cc-by."}, {"role": "assistant", "content": "The predictors can be classified as discrete or probabilistic depending on whether they provide a score, not necessarily a p value, for predictions. In the case of methods with discrete output, more or less ad hoc thresh- olds have been used to detect the most reliable events. Many machine learning based predictors are binary classi- fiers, however, it is possible to have more than two outputs e.g. by using multi-tier two-class prediction system. Features describe the characteristics of the investigated phenomenon. If several features are available it is impor- tant to choose those, which best capture the phenomenon. This is partly due to the curse of dimensionality, which means that much more data are needed when the number of features increases. The volume of the feature space grows exponentially with the dimensionality such that the data become sparse and insufficient to adequately describe the pattern in the feature space. Another problem is over- fitting, which means that the learner, due to sparse data, complex model or excessive learning procedure, describes noise or random features in the training dataset, instead of the real phenomenon. It is crucial to avoid overfitting as it leads to decreased performance on real cases. Many predictors provide a measure for the probability of prediction, in this domain a measure of how likely the var- iation is pathogenic. This information can be used for ranking the investigated cases. A more advanced version is to obtain e.g. by bootstrapping an estimate of the standard error of the prediction indicative of the prediction reliability. Many types of biological data are limited in quantity. The same data cannot be used both for method training and testing. The trick is to partition the dataset. This can be done in different ways, with cross-validation probably being the most popular of these. The dataset is divided into k disjoint partitions, one of which is used for testing and the others for training. This is repeated k times until all the partitions have been used as test set. Ten parti- tions i.e. ten times cross validation is the most common partitioning scheme. The average performance measures computed from the splits are used to describe the overall prediction performance. Random sampling is another approach, however, a problem is that the same cases may appear more than once in the test set and others not at all. Another computationally intensive validation approach is leave one out validation, an extreme case of cross validation with partitioning to the total number of instances. As the name implies, one case at time is left for validation while the remaining cases are used for training. The computational requirements may be prohi- bitive with large datasets. A problem especially for the last scheme is if there are some very similar cases in the dataset. Typically the training set should contain about equal amount of cases in each class. Imbalance in the numbers of cases in the classes can cause problems during perfor- mance evaluation as discussed below. There are some ways to handle class imbalance. Principles of method evaluation To test and compare predictors two requirements have to be met. There has to be available test dataset with known outcome and there has to be in place suitable pre- diction performance evaluation measures. Benchmark is a gold standard dataset - cases with experimentally vali- dated known effects which represent the real world. These can be used for training machine learning methods as well as for testing the developed methods. The same Figure 2 Principles of machine learning Machine learning is a form of supervised learning in which a computer system learns from given positive and negative instances to distinguish between cases belonging to the two classes. During training, positive and negative cases (black and white balls) are provided for the system, which leads to organization of the predictor (indicate by the arrangement of the black and white squares inside the predictor) such that it learns to separate the cases and thus can classify unknown cases (balls with question marks). Depending on the classifier, whether it yields in addition to the classification also a score for the prediction, the results can be called as discrete or probabilistic."}], "source_title": "How to evaluate performance of prediction methods? Measures and their interpretation in variation effect analysis", "source_doi": "10.1186/1471-2164-13-s4-s2", "source_url": "https://bmcgenomics.biomedcentral.com/counter/pdf/10.1186/1471-2164-13-S4-S2", "source_license": "cc-by", "zotero_item_key": "LKJJ9WPZ"}
{"id": "34ba43f69e1a0fda", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Reinforcement learning based process optimization and strategy development in conventional tunneling; DOI: 10.1016/j.autcon.2021.103701; license: cc-by."}, {"role": "assistant", "content": "Lanham, Reinforcement Learning, from Games to Geologic Interpretation. htt ps://medium.com/@cxbxmxcx/reinforcement-learning-from-games-to-geologi c-interpretation-93757664f0e4, 2019 (accessed 2 October 2020). [18] M. Lanham, Reinforcement Learning, It’s Coming not Just for Games. https://me dium.com/@cxbxmxcx/reinforcement-learning-its-coming-not-just-for-games-6b6 064e02bbe, 2019. [19] D. Schwung, A. Schwung, S. Ding, Actor-critic reinforcement learning for energy optimization in hybrid production environments, Int. J. Comput. 18 (2019) 360–371. [20] I. Carlucho, M. de Paula, S. Wang, Y. Petillot, G.G. Acosta, Adaptive low-level control of autonomous underwater vehicles using deep reinforcement learning, Robot. Auton. Syst. 107 (2018) 71–86, https://doi.org/10.1016/j. robot.2018.05.016. [21] B. Maidl, M. Thewes, U. Maidl, D. Sturge, Handbook of Tunnel Engineering, First, Engl. ed., Ernst/Wiley, Berlin, 2013. [22] W. Stipek, R. Galler, M. Bauer (Eds.), 50 Years of NATM: Experience Reports, ITA, Austria, Wien, 2012. [23] Norwegian Tunnelling Society (Ed.), Norwegian Tunnelling Technology: Publication no. 23, 2014. [24] S. Pelizza, D. Peila, Soil and rock reinforcements in tunnelling, Tunn. Undergr. Space Technol. 8 (1993) 357–372, https://doi.org/10.1016/0886-7798(93)90020- V. [25] P. Lunardi, Design and Construction of Tunnels: Analysis of Controlled Deformation in Rocks and Soils (ADECO-RS), Springer-Verlag, Berlin, Heidelberg, 2008. [26] S. Raschka, V. Mirjalili, Python Machine Learning - Third Edition: Machine learning and deep learning with python, scikit … -learn, and tensorflow 2, Packt Publishing Limited, [S.l.], 2019. [27] R.S. Sutton, A.G. Barto, Reinforcement Learning: An Introduction, Second edition, The MIT Press, Cambridge Massachusetts, 2018. [28] ITA Working Group Conventional Tunnelling, General Report on Conventional Tunnelling Method, Distrib, ITA Secretariat c/o EPFL, Lausanne, 2009. [29] P.A. Vermeer, N. Ruse, T. Marcher, Tunnel heading stability in drained ground, Felsbau 20 (2002) 8–18. [30] M.A. Kanji, Critical issues in soft rocks, J. Rock Mech. Geotech. Eng. 6 (2014) 186–195, https://doi.org/10.1016/j.jrmge.2014.04.002. [31] T. Marcher, S. Stauder, M. Winkler, HSSR – Ein Versuch der Einordnung und Abgrenzung des Materials, in: T. Marcher (Ed.), Beitr¨age zum 1. Hard Soil – Soft Rock (HSSR) Minisymposium: Charakterisierung, Modellierung und experimentelle Untersuchungen von Uebergangsgesteinen – Aktuelles aus Forschung und Entwicklung, Graz, 2020. [32] G. Anagnostou, K. Kov´ari, Face stability conditions with earth-pressure-balanced shields, Tunn. Undergr. Space Technol. 11 (1996) 165–173, https://doi.org/ 10.1016/0886-7798(96)00017-X. [33] A.A. Borovkov, Random walks and factorisation identities, in: A.A. Borovkov (Ed.), Probability Theory, Springer London, London, 2013. [34] C.J.C.H. Watkins, Learning from Delayed Rewards. Ph.D. Thesis, Oxford, 1989. [35] C.J.C.H. Watkins, P. Dayan, Q-learning, Mach. Learn. 8 (1992) 279–292, https:// doi.org/10.1007/BF00992698. [36] A.A. Javadi, M. Rezania, Applications of artificial intelligence and data mining techniques in soil modeling, Geomech. Eng. 1 (2009) 53–74, https://doi.org/ 10.12989/gae.2009.1.1.053. [37] M.A. Shahin, M.B. Jaksa, H.R. Maier, Recent advances and future challenges for artificial neural Systems in Geotechnical Engineering Applications, Adv. Artific. Neural Syst. 2009 (2009) 1–9, https://doi.org/10.1155/2009/308239. [38] H. Kinsley, Reinforcement Learning W/ Python. https://pythonprogramming.net /q-learning-reinforcement-learning-python-tutorial/, 2019. [39] R.H.R. Hahnloser, R. Sarpeshkar, M.A. Mahowald, R.J. Douglas, H.S. Seung, Digital selection and analogue amplification coexist in a cortex-inspired silicon circuit, Nature 405 (2000) 947–951, https://doi.org/10.1038/35016072. [40] X. Glorot, A. Bordes, Y. Bengio, Deep sparse rectifier neural networks, in: G. Gordon, D. Dunson, M. Dudik (Eds.), Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics, 2011, pp. 315–323. [41] Y. Lecun, L. Bottou, Y. Bengio, P. Haffner, Gradient-based learning applied to document recognition, Proc. IEEE 86 (1998) 2278–2324, https://doi.org/10.1109/ 5.726791. [42] D.H. HUBEL, T.N. WIESEL, Shape and arrangement of columns in cat’s striate cortex, J. Physiol. 165 (1963) 559–568, https://doi.org/10.1113/jphysiol.1963. sp007079. [43] F."}], "source_title": "Reinforcement learning based process optimization and strategy development in conventional tunneling", "source_doi": "10.1016/j.autcon.2021.103701", "source_url": "https://www.sciencedirect.com/science/article/pii/S0926580521001527", "source_license": "cc-by", "zotero_item_key": "N5Q33JID"}
{"id": "07163f3041d054a0", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Reinforcement learning based process optimization and strategy development in conventional tunneling; DOI: 10.1016/j.autcon.2021.103701; license: cc-by."}, {"role": "assistant", "content": "In this paper we present a novel RL based framework for construction process optimization and strategy development for conventional * Corresponding author. E-mail address: erharter@tugraz.at (G.H. Erharter). Contents lists available at ScienceDirect Automation in Construction journal homepage: www.elsevier.com/locate/autcon https://doi.org/10.1016/j.autcon.2021.103701 Received 1 December 2020; Received in revised form 29 March 2021; Accepted 2 April 2021\n[page 2] Automation in Construction 127 (2021) 103701 2 tunneling. Such models can act as decision support for the geotechnical engineer, engineering geologist, geotechnician etc. (hereafter “geo­ technician” is used) (design choices, progress-planning) and in the long run such models work towards full automation in underground con­ struction. Hence, the model is a first attempt to automate decisions made by the geotechnician on face in underground construction. In the next section (2) we frame the process of conventional tunneling as a RL problem and provide details on how these two disci­ plines can connect. Section 3 is the main methodological section that presents the geotechnical scenario at the background of this RL simu­ lation as well as the agent and the environment. In Section 4 we describe the training process and in Section 5 we show experiences gathered during the training and testing of agents. A conclusion and implications for the vision of “digital tunneling” is given in Section 6 and we present an outlook in the last Section 7. A reference to the Python based code for this paper, is given in the appendix. 2. Conventional tunneling as a reinforcement learning problem The process of reinforcement learning (RL) is typically depicted as a closed loop where an agent takes different actions, to influence an environment which responds by sending an updated state as well as a reward signal to the agent (e.g. [26]). To apply RL, the learning problem must undergo the Markov property, i.e. we only need to know the cur­ rent state of the system, to make a decision [27]. Therefore, the state must include information about all aspects of the past agent–environ­ ment interaction that make a difference for the future. This can be said to be true for tunnel excavation where we only need to know the state of the rockmass and the excavation-process to decide on how to proceed. Tunnel construction follows several cyclic and sequential processes, some of which can be framed as loops and therefore translated to RL problems. The most outstanding loop in conventional tunneling is the excavation of an underground opening with a sequential construction process of: blasting, mucking and rock support installation [28] (“excavation loop” in Fig. 1). While the components of this excavation loop are themselves often sequential processes (e.g. blasting sequence, support installation etc.), the excavation loop specifically is based on a sequence of geotechnical decisions belonging to a bigger cycle which we refer to as the excavation sequence decision – loop. Looking at the bigger picture, excavation sequence decisions are one part of the whole con­ struction phase of a tunnel and therefore part of the whole tunnel life cycle (Fig. 1). The focus of this study is to create a simulation of a simplified version of the process that governs the general excavation sequence decisions and frame it as a RL-loop. This process can be translated to a RL-loop/ Markov decision process [27] consisting of the following components: ▪ the decision making geotechnician is the agent ▪ processes like “top heading excavation”, “bench excavation”, “installation of face support” etc. are the actions ▪the rockmass itself and the construction site with all its pro­ cesses are the environment (here described at each timestep by the state of the environment and the reward-system) ▪the sum of all delays (planned and unplanned) and complica­ tions throughout the course of the excavation which result from the geotechnician’s actions are the reward ▪ the current state of the excavation including information about the past and recent rockmass conditions as well as the already installed support are the state Fig. 2 is a graphical representation of this process which we refer to as “TunnRL” (Tunnel automation with Reinforcement Learning). The individual components of this loop in the above given list as well as in Fig. 1 and Fig. 2 are only for explanatory purpose and do not claim to be complete."}], "source_title": "Reinforcement learning based process optimization and strategy development in conventional tunneling", "source_doi": "10.1016/j.autcon.2021.103701", "source_url": "https://www.sciencedirect.com/science/article/pii/S0926580521001527", "source_license": "cc-by", "zotero_item_key": "N5Q33JID"}
{"id": "6e465eb26aa5bf7d", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Cutterhead and Cutting Tools Configurations in Coarse Grain Soils; DOI: 10.2174/1874836801711010182; license: cc-by."}, {"role": "assistant", "content": "When the plate type cutterhead is employed for shield machines in the coarse grain ground, disc cutters in combination with soft ground cutting tools (scraper, ripper and pick) are usually mounted on the cutterhead, and the installed disc cutters are tools to break big cobble/boulders into pieces, and the remaining work is fulfilled by the soft ground tools. 3.2. Performances of Shield Cutterheads The two types of cutterheads are both adopted in China subway/metro construction to dig tunnels in the coarse grain soils, but received different responses. Effects of the cutterheads equipped with the corresponding cutting tools are revealed and illustrated by the following construction cases. 3.2.1. Cases of Using the Plate Type Cutterhead (1) The plate type cutterhead EPB shield was used in some project of Beijing subway Line 5. After driving 346 m, the machine slowed down and a sharp increase in the torque of cutterhead took place. Through looking over the cutterhead with access to the plenum (excavation chamber), it was observed that the four pairs of scrapers at the outer side of the cutterhead were totally worn out and the face scrapers were also seriously abraded (see Fig. (4)) and this was (a) Cutterhead with disc cutters. (b) Cutterhead without disc cutters.\n[page 7] Cutterhead and Cutting Tools Configurations The Open Construction and Building Technology Journal, 2017, Volume 11 187 also true of the main body of the cutterhead. After restoring the cutterhead with repair welding, reinstalling all side scrapers and replacing 68 face scrapers, the shield machine started to work again. The main reasons for the excessive wears are as follows: the small opening ratio less than 30%; the four-spoke design and not enough scrapers; large forces acting on the scrapers on account of not installing precut bits and center fish-tail bit. Fig. (4). Excessive wears of the cutterhead used for some project of Beijing subway Line 5. (2) A running tunnel on Beijing subway Line 4 was bored by an EPB shield machine equipped with disc cutters. The construction work was completed after many times of replacements of disc cutters. It was observed from Fig. (5) that the cutterhead and the cutting tools were seriously damaged, all top layers of the breasting plate located out of the central area of the cutterhead fell off, and the space between the remaining middle and bottom layers of the plate was full of muck. The causes for the damages can be summarized as: the small opening ratio of about 25% of the cutterhead; no enough reaction forces on disc cutters resulting in abnormal wears of disc cutters; not installing precut bits including nose cone bits and excessive secondary wears by rolling of the over-sized grains between the breasting plates and the cutting face. Fig. (5). Damaged cutterhead of some project of Beijing subway Line 4. (3) Two EPB shield machines with plate type cutterheads were adopted in building running tunnels of some project of Chengdu metro line 2. The opening ratio of cutterhead is about 20% and not equipped with precut bits, as shown in Fig. (6). Due to the large content of over-sized particles, the abnormal wears of the disc cutters were serious and frequent replacements of tools were performed per 50-100 m. After changing types of the alloys and workmanship for the scrapers, installing new scrapers in some disc cutter boxes, and welding the reinforced-type precut bits to the cutterhead, the boring length before the next replacing tools was increased to more than 300 m. The reasons for the tough driving work are as follows: small opening ratio of about 20%; unreasonable configurations of cutting tools and cutterhead. The driving distance was only prolonged by the improved cutting tools because of the unmodifiable structure of the cutterhead.\n[page 8] 188 The Open Construction and Building Technology Journal, 2017, Volume 11 Li et al. Fig. (6). Cutterhead and damaged disc cutters of some project of Chengdu metro Line 2. 3.2.2. Cases of Using the Spoke Type Cutterhead The spoke type cutterhead was ever widely used in Beijing shield tunneling. In the following construction cases the shield machines of 6.14 m diameter are employed, and the cutterheads are given in (Fig. 7). The six cylindrical spoke type cutterhead with opening ratio of 75.5% is equipped with 82 scrapers, 28 precut bits (ripper teeth), 12 shell bits, one center fish-tail bit and two copy cutters, as presented in (Figs. 8-11). Fig. (7)."}], "source_title": "Cutterhead and Cutting Tools Configurations in Coarse Grain Soils", "source_doi": "10.2174/1874836801711010182", "source_url": "https://openconstructionbuildingtechnologyjournal.com/VOLUME/11/PAGE/182/PDF/", "source_license": "cc-by", "zotero_item_key": "7KDIA2X2"}
{"id": "d6caa4f044930b7b", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: How to evaluate performance of prediction methods? Measures and their interpretation in variation effect analysis; DOI: 10.1186/1471-2164-13-s4-s2; license: cc-by."}, {"role": "assistant", "content": "Hamming distance and quadratic distance (also called for Euclidean distance), which are the same for binary data, and relative entropy and mutual information [27]. ROC analysis Receiver operating characteristics (ROC) analysis is a visualization of prediction performance, that can be used to select suitable classifier (for review see [28,29]). It indi- cates the tradeoffs between sensitivity and specificity. ROC curves can be drawn with specific programs when the pre- dictor is of probabilistic type and provides a score for the classification. The score is usually not a real p value, but a value usable for ranking the predictions. ROC curve (Fig. 7a) is drawn by first ranking the data based on the prediction score. Then the data are divided to intervals of equal size. The upper limit for the partitions is the number of cases in the dataset. ROC curve has on x-axis 1-specificity also called FPR and on the y-axis sensi- tivity (TPR). Computer program establishes cut offs at intervals, calculates contingency table for data in the interval, and determines the values for sensitivity and 1-specificity, which is plotted to the graph. The procedure is repeated for each partition. If cross validation has been used, then the ROC curve can be used to show the average and variance of the results. In an ideal case all the true positive cases are on the first half of the ranked list and the plot rises to (0,1) and then continues straight to the right with all the true negative cases. A random classification would be on the diagonal i.e. mixed correct and wrong cases. The faster the curve rises and the higher it reaches in the beginning the better the method is. Methods can be compared with ROC ana- lysis when the same test dataset (benchmark) is used (Fig 7b). The curve that runs higher is for a better method. If the curves cross (Fig 7c) the comparison is no more meaningful. Area under the ROC curve (AUC) has been used as a measure of goodness for predictions (Fig. 7a). It approxi- mates the probability of ranking a randomly chosen posi- tive instance higher than a randomly chosen negative one. A value of 0.5 indicates random and useless classification while 1 would indicate perfect classifier. Note that AUC Figure 7 ROC analysis and AUC a) Principle of ROC analysis. b) Comparison of predictors based on the ROC curves when the methods are tested with the same dataset (benchmark). c) If the curves cross the comparison is no more meaningful. Vihinen BMC Genomics 2012, 13(Suppl 4):S2 http://www.biomedcentral.com/1471-2164/13/S4/S2 Page 7 of 10\n[page 8] can be even smaller than 0.5. One should bear in mind that the ROC curve does not directly indicate the perfor- mance of a method. It shows the method’s ranking poten- tial, which is related to overall performance, further strengthening the fact that a single measure cannot fully describe the predictive performance even if it produces a graph. What if the data is classified to more than two classes? If there are more than two classes the measures described above cannot be applied. The data can still be presented in an N x N contingency table. One approach is to divide the data into several partitions of two categories. If parameters are needed for all the classes there are some options available, however, single measures are more problematic. It is possible to calculate row and col- umn wise ratios in the same way as in Fig. 4. MCC is in fact a special case for binary data of linear correlation coefficient, which can be used for several classes in its general format. Mutual information analysis can be used in these cases, as well. Applicable measures have been discussed e.g. in [27]. Examples of performance comparisons This section discusses examples of variation effect pre- diction method evaluations. These include methods for amino acid substitution (missense variation) tolerance, point variation effects on protein stability and variations related to mRNA splicing. The discussion concentrates on the comparison principles, especially in the light of the discussion on requirements mentioned above. The actual comparisons are not presented as it would have required publication of substantial parts of the reports. As a single parameter is insufficient for ranking meth- ods, the readers are directed to the original articles to find all the details. Here a summary to the methodology and use of the evaluation parameters is provided."}], "source_title": "How to evaluate performance of prediction methods? Measures and their interpretation in variation effect analysis", "source_doi": "10.1186/1471-2164-13-s4-s2", "source_url": "https://bmcgenomics.biomedcentral.com/counter/pdf/10.1186/1471-2164-13-S4-S2", "source_license": "cc-by", "zotero_item_key": "LKJJ9WPZ"}
{"id": "4343e828823e3756", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Reinforcement learning based process optimization and strategy development in conventional tunneling; DOI: 10.1016/j.autcon.2021.103701; license: cc-by."}, {"role": "assistant", "content": "As the present paper should be the first introduction of RL into tunneling we refrained from over-complicating the simulation and rather give an impetus for future studies that build upon this work. The below given improvements are a non-exhaustive list of ideas that we think are worth to be further explored. Ever improving processing power will alleviate problems related to computational cost over time. Nevertheless, all improvements must consider that each step in the framework will be done millions of times throughout the training and state of the art RL is by itself already computationally heavy. Improvements address either the agent or the environment: ▪ On the environment’s side, improvements could work towards making the geotechnical scenario more realistic by involving more ground types and other phenomena like groundwater and in-situ stress conditions. The excavation geometry could become more complex so that also other excavation shapes and sequences such as full-face excavation, or a further division of the top heading and bench excavation are possible. ▪ Closely related to the excavation geometry – and in our view one of the main points that should be improved – is the stability evaluation of the simulation. The reason why we used Eq. (3) after [29] as a substitute for more sophisticated means of sta­ bility assessment (e.g. tunnel cross sectional analyses such as analytical convergence confinement methods or 2D and 3D finite element analysis) is that this analytical solutions is computationally very efficient and does not prolongate the training process too much. For example, 100,000 episodes of the current framework with around 120 moves per episode would require ~1.2e+7 FEM based stability assessments. However, this reduction of the stability assessment is only us­ able for the ground conditions described in [29] and also ne­ glects other phenomena like the 3D stress state at the tunnel face. Improvements could be to do the stability assessment based on stand-up time concepts that also involve the rockmass quality [50] or ideally 3D finite element analysis as given above. ▪ Improvements for the agent on the one hand address the agent’s performance by modifying the agent’s architecture itself. Sys­ tematic hyperparameter tuning was not conducted yet and may Fig. 9. Histograms showing the performance of the agent 4 that was tested for 10,000 episodes in Fig. 7. Note the different scales of the x-axes in comparison to the histograms of Fig. 8. Fig. 10. Boxplot that shows how many times the agent 4 uses each action throughout 10,000 test episodes. Bold black lines in the boxes represent median values; the boxes confine the upper and lower quartiles and the whiskers show min.-max. Values; “al” in the x-labels refers to “advance length”. G.H. Erharter et al.\n[page 11] Automation in Construction 127 (2021) 103701 11 help to further improve the agent’s performance (e.g. grid−/ random search or even RL based hyperparameter optimization [51]). From a geotechnical point of view though, interesting improvements mainly concern an extension of the agent’s ca­ pabilities by introducing more possible actions. Whereas direct improvements to the given framework would be more advance lengths and types of tunnel support, other ideas are to involve exploration ahead of the face by simulated measurement while drilling [52] or geophysical exploration [53] to give the agent an idea what might be in front of the current excavation face. An idea in this regard is also to extend the agent to a multi- agent framework as it was used successfully before [15] where different agents have different tasks to fulfill. Future studies will work towards a more realistic environment and more complex agents in the TunnRL framework. Where TunnRL fits well in the line of the current development of automation, the greatest po­ tential lies in the possibility to develop new and not yet considered tunneling strategies for sequential tunnel excavation. Declaration of Competing Interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Appendix Appendix 1: Link to the GitHub repository where this paper’s code can be found: https://github.com/geograz/Tunnel-automation-with-Re inforcement-Learning-TunnRL- References [1] T. Marcher, G.H. Erharter, M."}], "source_title": "Reinforcement learning based process optimization and strategy development in conventional tunneling", "source_doi": "10.1016/j.autcon.2021.103701", "source_url": "https://www.sciencedirect.com/science/article/pii/S0926580521001527", "source_license": "cc-by", "zotero_item_key": "N5Q33JID"}
{"id": "cf1cfae89bdd952b", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Cutterhead and Cutting Tools Configurations in Coarse Grain Soils; DOI: 10.2174/1874836801711010182; license: cc-by."}, {"role": "assistant", "content": "As shown in Fig. (15), the scrapers were seriously collided and part tip of the center fish-tail bit was broken off and the two sides of the ripper tooth were damaged though it remained good shape and could still work. Fig. (15). Damages of Fish-rail and precut bit caused by collisions. 3.2.3. Cases of Using the Spoke-plate Type Cutterhead (1) The running tunnels of Lot 4 of Beijing subway Line 4 were bored using two shield machines equipped with the spoke-plate type cutterhead of 39% opening ratio, as shown in (Fig. 16). The ground is very much alike that encountered at Lot 14 of Beijing subway Line 4 mentioned above. The combination of cutting tools is very similar to that used in Lot 14 of Beijing subway Line 4, consisting of 94 scrapers, 40 precut bits, 12 shell bits, one center fish-tail bit and two copy cutters. The running tunnels were finished without changing tools. The cutterhead and some tools after driving are presented in (Fig. 17). (a) Damaged fish-tail bit. (b) Damaged precut bit.\n[page 13] Cutterhead and Cutting Tools Configurations The Open Construction and Building Technology Journal, 2017, Volume 11 193 Fig. (16). Spoke-plate type cutterhead. Fig. (17). Used cutterhead and the damaged cutting tools. (a) Used cutterhead. (b) Damaged precut bits. (c) Damaged scrapers.\n[page 14] 194 The Open Construction and Building Technology Journal, 2017, Volume 11 Li et al. (2) As shown in Fig. (18), the spoke-plate type cutterhead with opening ratio of 27% was used in building running tunnels of Lot 9 of Shenyang metro Line 1. The tunnels mainly passed through the sandy gravel/cobble soils and the maximum size of particle was about 200 mm, and the initial cutting tools were face scrapers and circumferential bucket scrapers. After driving the No. 193 ring, the cutterhead and cutting tools were seriously destroyed. Main retrofitting jobs included wear resistant plating for cutterhead, installing 28 precut bits as well as repairing damaged scrapers and tooth bits (see Fig. (19)). Using the improved cutterhead and cutting tools, the remaining part of the tunnel was bored without interruption. Fig. (18). Initial cutterhead and the cutting tools. Fig. (19). Retrofitted cutterhead and cutting tools. 4. DISCUSSIONS AND SUGGESTIONS Though configurations of the cutterhead and tools of a shield machine must be customized for specific ground conditions, some options with respect to key parameters of the configurations can be deduced from the above- mentioned construction cases. 4.1. Opening Ratio of Cutterhead Although the opening ratio of a cutterhead is the result of a compromise involving a large number of aspects, a high opening ratio of the shield cutterhead used in sandy gravel/cobble soils with boulders is a better choice, as has been proved by the many above cases of using the spoke type and spoke-plate type cutterheads. In these cases, the tunnels were bored without change cutting tools. Whereas in the cases of using the plate type cutterheads with low opening \u000e \u000f \b \u0005\u0006\u0004\u0018\u0005 \u000f \u0004\u001a\u0004\u0018 \u0019 \u0011\u0010 \u0004 \u0006 \u0010\u000f \u0015\u0004\u000f \u001a\u0015\u000f\u0013!\u0004\"\u0006 \u0013\u0013 \u0003\u0006\u0007\b \b \u0006\u0015\u000f\u0004\u0010\u0011 \u0006\u0007\b \u0002\u0010\u000f \u0015\u0004\u000f \u000e\u000f\u0004\u0010\u0011 \u0006\u0007\b \u0014\u0013\u0013 \u0015\u0013\u0013 \u0016\u0017\u0013\u0013\u0010\u0013\u0013\u0013 \u0018\u000f\u0013\u0013\u0019\u0013\u0013\u0013\u0013\u0013 \u0018\u001a\u0013\u0013\u0019\u0013\u0013\u0013\u0013\u0013 \u0015\u0013\u0013 \u0014\u0013\u0013 \u0010\u0013\u0013\u0019\u0013\u0013\u0013\u0013\u0013 \u0018\u000f\u0013\u0013\u0019\u0013\u0013\u0013\u0013 \u0010\u001a\u0012\u0013\u0013\u0019\u0013\u0013\u0013 \u0015\u0013\u0013 \u0018\u0017\u0013\u0013\u0019\u0013\u0013 \u0018\u0015\u0013\u0019\u0013\u0013\u0013\u0013 \u001a\u0013\u0013\n[page 15] Cutterhead and Cutting Tools Configurations The Open Construction and Building Technology Journal, 2017, Volume 11 195 ratios of 20%-30%, the interventions for replacing cutting tools are indispensable to realizing driving of the whole tunnel at the expenses of additional time and money. The spoke type and spoke-plate type cutterheads with high opening ratios of 39-75.5% are favorable to passing of the ripped-off big stones into the rear mixing chamber without the need of a substantial pressure drop, thus mitigating wears and damages of cutting tools and other components as well as reducing interventions of changing tools. This is one of the main reasons why the spoke type cutterhead is widely adopted in Beijing to construct subway tunnels. The optimal opening ratio of a cutterhead exists, as can be deduced from the calculated wear coefficients of cutting tools in the projects of Lot 4 and 14 of Beijing subway Line 4. The wear coefficient is defined as the wear amount dividing by the tool path length. The involved strata are almost identical and configurations of the cutting tools are very alike, as previously mentioned. The average wear coefficients for precut bits, including shell bits at the periphery, of the two cutterheads are about 0.035 mm/km."}], "source_title": "Cutterhead and Cutting Tools Configurations in Coarse Grain Soils", "source_doi": "10.2174/1874836801711010182", "source_url": "https://openconstructionbuildingtechnologyjournal.com/VOLUME/11/PAGE/182/PDF/", "source_license": "cc-by", "zotero_item_key": "7KDIA2X2"}
{"id": "865e54f5143cd2ef", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Intelligent Classification of Surrounding Rock of Tunnel Based on 10 Machine Learning Algorithms; DOI: 10.3390/app12052656; license: cc-by-sa."}, {"role": "assistant", "content": "The average accuracy was calculated by first averaging the accuracy of each validation set over the three surrounding rock grades, and then averaging the averages over the five validation sets. In the final evaluation of the model performance, the precision, recall, F1 score, and accuracy were adopted for a comprehensive evaluation. And these indexes are the average of the results of the three test sets. The process is shown in Figure 8. Figure 8. Model parameter optimization and evaluation process. The performance of the models after training is shown in Figure 9 and Table 7.\n[page 13] Appl. Sci. 2022, 12, 2656 13 of 20 (a) (b) (c) (d) Figure 9. Model parameter optimization and evaluation process: (a) precision; (b) recall; (c) F1; (d) accuracy. Table 7. Performance of the intelligent surrounding rock models. Performance Index Rock Grade Supervised Learning Algorithm Average SVM BP RBF KNN GNB DT RT ET Bag-ging GB Precision III 0.81 0.79 0.73 0.77 0.75 0.75 0.81 0.80 0.80 0.81 0.78 IV 0.90 0.87 0.76 0.84 0.75 0.59 0.85 0.88 0.83 0.87 0.81 V 0.86 0.88 0.87 0.90 0.82 0.84 0.90 0.92 0.91 0.88 0.88 Recall III 0.96 0.97 0.94 0.93 0.91 0.73 0.92 0.94 0.93 0.96 0.92 IV 0.62 0.61 0.52 0.63 0.50 0.60 0.67 0.67 0.67 0.65 0.61 V 0.97 0.94 0.89 0.94 0.92 0.86 0.96 0.96 0.93 0.94 0.93 F1 score III 0.88 0.87 0.82 0.84 0.82 0.74 0.86 0.87 0.86 0.88 0.84 IV 0.73 0.72 0.62 0.72 0.60 0.59 0.75 0.76 0.74 0.75 0.70 V 0.91 0.91 0.88 0.92 0.87 0.85 0.93 0.94 0.92 0.91 0.90 Accuracy - 0.85 0.84 0.78 0.84 0.78 0.73 0.85 0.86 0.84 0.85 0.82 The following conclusions can be drawn from Figure 9 and Table 7. Among 10 machine learning algorithm models, the average precision, recall, F1 score and accuracy were above 0.7, except for the recall of grade-IV surrounding rock. In particular, the average recalls of grade-III and -V were greater than 0.90. Thus, these machine learning models established by the drilling parameters are feasible and reliable in the intelligent classification of surrounding rocks.\n[page 14] Appl. Sci. 2022, 12, 2656 14 of 20 2.8. Model Selection and Recommendation When comprehensively considering the precision, recall, F1 score, and accuracy, three types of models (SVM, RT, and ET) had better performance among the 10 machine learning algorithm models. Their average precision, recall, F1 score, and accuracy were above 0.8, except for the recall of grade-IV surrounding rock. In particular, the recall of the grade-V surrounding rock, which has a great influence on the safety of the tunnel, was above 0.95. Thus, these machine learning models based on SVM, RT, and ET have some degree of safety and high practical value. 3. Case Study 3.1. Intelligent Surrounding Rock Classification System Software The tunnel surrounding rock intelligent classification system by the drill and blast method was established based on the intelligent surrounding rock classification models based on SVM, RT, and ET. It can automatically record and transmit the drilling parameters and intelligently classify the surrounding rock with carriers of the intelligent drill jumbo. Specifically, the surrounding rock grade of the tunnel face is determined by the votes of three models (SVM, RT, and ET). When all types of the surrounding rock levels are not superior, the higher grade is taken as the surrounding rock grade of the tunnel face in consideration of safety. In addition, these samples for the models do not require consideration of the modifica- tion conditions (i.e., low crustal stress, without groundwater or drip outflow of ground- water, and without a main weak structure surface). Therefore, in consideration of safety and popularization, the grade determined from these models is the basic surrounding rock grade; the grade amended by the groundwater state, major weak structural surface, and crustal stress state is the final surrounding rock grade. The specific process is shown in Figure 10. Figure 10. Classification process of the intelligent surrounding rock classification system. The system interface is shown in Figure 11.\n[page 15] Appl. Sci. 2022, 12, 2656 15 of 20 grade-VI grade-V grade-IV grade-III grade-II grade-I Detials Figure 11. Interface of the ‘tunnel surrounding rock intelligent classification system by the drill and blast method’. 3.2. Project Overview of Test Tunnel The Gaojiaping Tunnel of the Zhengzhou–Wanzhou high-speed railway is located in Nanzhang county, Xiangyang city, Hubei province, China."}], "source_title": "Intelligent Classification of Surrounding Rock of Tunnel Based on 10 Machine Learning Algorithms", "source_doi": "10.3390/app12052656", "source_url": "https://www.mdpi.com/2076-3417/12/5/2656/pdf?version=1646383280", "source_license": "cc-by-sa", "zotero_item_key": "UGAERXYU"}
{"id": "7f414c3252e4f322", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Reinforcement learning based process optimization and strategy development in conventional tunneling; DOI: 10.1016/j.autcon.2021.103701; license: cc-by."}, {"role": "assistant", "content": "A solution would be to take a version of the agent that was saved before instability occurred and continue training with that, but we left Fig. 7. Different training paths for five exemplary agents over 120,000 episodes. The first row shows cumulative rewards per episode; the second row the average number of blasts/moves that were required to complete each episode; the third row the average number of face instabilities per episode; the fourth row shows the average loss per episode. Transparent colors in the background show the raw data records and the solid lines in the foreground a 500-episode sliding window average. Fig. 8. Histogram for the performance of an agent who plays 10,000 episodes with completely random moves (i.e. ε = 1). G.H. Erharter et al.\n[page 9] Automation in Construction 127 (2021) 103701 9 the agents 2 and 5 as they are for explanatory purposes. Before discussing the performance of the agents, we first establish a baseline for the minimum performance that an agent must reach by letting an agent play with completely random moves (i.e. ε = 1) for 10,000 episodes. In these 10,000 random episodes the agent reached a ▪ minimum/maximum/median reward of −544/255/-13 points, ▪ minimum/maximum/median number of 0/27/8 instabilities per episode, ▪ minimum/maximum/median number of 124/197/145 blasts/ moves per episode. Histograms for the 10,000 random episodes are given in Fig. 8. The maximum rewards, minimum number of face instabilities and minimum number of blasts/moves that the five agents of Fig. 7 needed are given in Table 5. We computed these numbers based on sliding window averages of 500 episodes as to avoid individual episodes that performed extraordinarily well. Albeit substantial differences can be seen in these statistics, it can generally be observed that all maximum rewards of the agents are above the maximum rewards of the randomly played episodes and therefore training was generally a success. In these five agents, two groups can be observed, where the agents 2 and 3 reached their maximum performance between episodes 40,000 and 60,000 and the agents 1, 4 and 5 reached their maximum perfor­ mance between episodes 80,000 and 120,000. Although the first group has reached their peak performance sooner, the achieved reward is generally lower than that of the second group (see description of training paths above). We conclude this section by presenting the strategy that the best performing agent 4 has found after 119,000 episodes – at its peak per­ formance (in terms of highest reward and low instabilities). We tested the saved agent’s checkpoint for 10,000 episodes with a fixed ε of 0.05 which corresponds to the ε at that stage of training and is in accordance to [14] who recommend an ε > 0 also for testing, as this helps the agent to deal with unexpected situations. The histograms of Fig. 9 show the same test statistics as given for the random moves in Fig. 8. The strategy that the agent adopted is focused on long advance lengths without face support, in alternation with long advance lengths with face support. By doing so the agent avoids face instabilities, while also minimizing the required support (see Table 4 for the respective rewards/penalties). The boxplot of Fig. 10 illustrates this, as it can be observed that the majority of actions is 4 m long advance lengths. Furthermore, the agent focuses on long advance lengths without face support in both excavation types (top heading and bench) which shows that it tries to avoid excessive use of support measures, thus showing a tendency towards economical optimization. It can also be seen that the agent still uses small advance lengths sometimes and, in this case, favors the actions without face support as it has realized that the small advance lengths do not lead to face instabilities in the given conditions. In Fig. 11 an example of one episode for the agent 4 is visualized. The time-distance diagram in the top row of this figure shows that the agent has learned to optimize the excavation process by minimizing changes between top heading and bench excavation which would be associated with unwanted delays in “real life” tunneling (e.g. building and removing of access ramps). As given in Section 2.1, installed support ahead of the current face always covers 10 m. With the maximum advance length being set to 4 m it would be unnecessary and uneco­ nomical to install face support in two consecutive rounds. In the second row of Fig."}], "source_title": "Reinforcement learning based process optimization and strategy development in conventional tunneling", "source_doi": "10.1016/j.autcon.2021.103701", "source_url": "https://www.sciencedirect.com/science/article/pii/S0926580521001527", "source_license": "cc-by", "zotero_item_key": "RXHUVV3X"}
{"id": "44a3af44fb3f0b5c", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Application of machine learning in predicting the rate-dependent compressive strength of rocks; DOI: 10.1016/j.jrmge.2022.01.008; license: cc-by-sa."}, {"role": "assistant", "content": "Rock strength prediction in real-time while drilling employing random forest and functional network techniques. J. Energy Resour. Technol. 143 (9), 093004. Hochreiter, S., Younger, A.S., Conwell, P.R., 2001. In: Learning to Learn Using Gradient Descent. International Conference on Artificial Neural Networks. Springer, Berlin, Heidelberg. Huang, X., Qi, S., Xia, K., Shi, X., 2018. Particle crushing of a filled fracture during compression and its effect on stress wave propagation. J. Geophys. Res. Solid Earth 123 (7), 5559e5587. Hudson, J.A., 2012. Design methodology for the safety of underground rock engi- neering. J. Rock Mech. Geotech. Eng. 4 (3), 205e214. ISRM, 1979. Suggested methods for determining the uniaxial compressive strength and deformability of rock materials. Int. J. Rock Mech. Min. Sci. 16 (2), 135e140. Krauthammer, T., Elfahal, M.M., Lim, J., Ohno, T., Beppu, M., Markeset, G., 2003. Size effect for high strength concrete cylinders subjected to axial impact. Int. J. Impact Eng. 28 (9), 1001e1016. Li, Q.M., Lu, Y.B., Meng, H., 2009. Further investigation on the dynamic compressive strength enhancement of concrete-like materials based on split Hopkinson pressure bar tests. Part II: numerical simulations. Int. J. Impact Eng. 36 (12), 1335e1345. Li, W., Tan, Z., 2017. Research on rock strength prediction based on least squares support vector machine. Geotech. Geol. Eng. 35, 385e393. Li, X.B., Lok, T.S., Zhao, J., 2005. Dynamic characteristics of granite subjected to intermediate loading rate. Rock Mech. Rock Eng. 38 (1), 21e39. Li, Y., Zou, C., Berecibar, M., Nanini-Maury, E., Chan, J.C.W., van den Bossche, P., Mierlo, J.V., Omar, N., 2018. Random forest regression for online capacity esti- mation of lithium-ion batteries. Appl. Energy 232, 197e210. Masoumi, H., Saydam, S., Hagan, P.C., 2015. Unified size-effect law for intact rock. Int. J. GeoMech. 16 (2), 04015059. Munoz, H., Taheri, A., 2017. Specimen aspect ratio and progressive field strain development of sandstone under uniaxial compression by three-dimensional digital image correlation. J. Rock Mech. Geotech. Eng. 9 (4), 599e610. Rumelhart, D.E., Hinton, G.E., Williams, R.J., 1986. Learning representations by back propagating errors. Nature 323, 533e536. Samui, P., 2008. Support vector machine applied to settlement of shallow founda- tions on cohesionless soils. Comput. Geotech. 35, 419e427. Smola, A.J., Scholkopf, B., 2004. A tutorial on support vector regression. Stat. Comput. 14, 199e222. Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., Salakhutdinov, R., 2014. Dropout: a simple way to prevent neural networks from overfitting. J. Mach. Learn. Res. 15 (1), 1929e1958. Wang, Y., Tonon, F., 2011. Dynamic validation of a discrete element code in modeling rock fragmentation. Int. J. Rock Mech. Min. Sci. 48 (4), 535e545. Weng, L., Wu, Z., Liu, Q., 2020. Dynamic mechanical properties of dry and water- saturated siltstones under sub-zero temperatures. Rock Mech. Rock Eng. 53, 4381e4401. Whittington, W.R., Oppedal, A.L., Francis, D.K., Horstemeyer, M.F., 2015. A novel intermediate strain rate testing device: the serpentine transmitted bar. Int. J. Impact Eng. 81, 1e7. Xia, K., Nasseri, M.H.B., Mohanty, B., Lu, F., Chen, R., Luo, S.N., 2008. Effects of mi- crostructures on dynamic compression of Barre granite. Int. J. Rock Mech. Min. Sci. 45 (6), 879e887. Yılmaz, I., Yuksek, A., 2008. An example of artificial neural network (ANN) appli- cation for indirect estimation of rock parameters. Rock Mech. Rock Eng. 41, 781e795. Yu, M., Li, S., Sun, Q., Wang, S., 2021. Influence of grain size on the strain-rate- dependent dynamic response of sandstones. Geomech. Geophys. Geo. 7, 74. Zhang, Q.B., Zhao, J., 2013. Determination of mechanical properties and full-field strain measurements of rock material under dynamic loads. Int. J. Rock Mech. Min. Sci. 60, 423e439. Zhang, Z.X., Kou, S.Q., Jiang, L.G., Lindqvist, P.A., 2000. Effects of loading rate on rock fracture: fracture characteristics and energy partitioning. Int. J. Rock Mech. Min. Sci. 37, 745e762. Zhao, J., Zhou, Y.X., Hefny, A.M., Cai, J.G., Chen, S.G., Li, H.B., Liu, J.F., Jain, M., Foo, S.T., Seah, C.C., 1999. Rock dynamics research related to cavern development for ammunition storage. Tunn. Undergr. Space Technol. 14 (4), 513e526. Zhong, K., Zhao, W., Qin, C., Gao, H., Chen, W., 2021. Mechanical properties of roof rocks under superimposed static and dynamic loads with medium strain rates in coal mines."}], "source_title": "Application of machine learning in predicting the rate-dependent compressive strength of rocks", "source_doi": "10.1016/j.jrmge.2022.01.008", "source_url": "https://www.sciencedirect.com/science/article/pii/S167477552200049X", "source_license": "cc-by-sa", "zotero_item_key": "KNCWQQJD"}
{"id": "51e88ec909ed0da1", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Application of machine learning in predicting the rate-dependent compressive strength of rocks; DOI: 10.1016/j.jrmge.2022.01.008; license: cc-by-sa."}, {"role": "assistant", "content": "The strain rate is poorly correlated with the other parameters, as it is controlled by the loading condition. The spec- imen length and diameter are strongly correlated with each other but poorly correlated with the other parameters due to a length-to- diameter ratio of 0.5e1 suggested by the SHPB test standard (Zhou et al., 2012) and widely adopted in many studies. 2.2. Machine learning algorithms Three well-proved machine learning algorithms, i.e. SVM, BPNN, and RF, in rock mechanics were used. However, their applicability to rock dynamics has rarely been reported yet. Fig. 2 illustrates the specific steps of applying the three algorithms to predict the rate- dependent compressive strength of rocks. The basic features of these algorithms are summarized in Figs. 3e5. 2.2.1. SVM algorithm The SVM algorithm was originally proposed by Boser et al. (1992) for data classification and subsequently developed by Drucker et al. (1996) for regression and prediction analysis. The SVM based on the principle of structural risk minimization can minimize empirical risk and confidence range, as well as obtain excellent statistical laws even for small sample sizes (Samui, 2008). Fig. 3 shows a schematic diagram of the basic principle of the SVM to solve the regression and prediction problems. The goal of the SVM training is to find the best hyperplane closest to all data points. The training process is to continuously adjust the hyperplane pa- rameters to minimize the sum of the distances of all data points from the hyperplane plane. In this study, the SVM is used to construct the mapping relationship between the compressive strength and the other parameters in Table A1. Thus, the feature vector of the compressive strength can be written as {xi, yi} (i ¼ 1, 2, ., m; xi ˛ Rd; yi ˛ R), where x represents the input, y denotes the output, m is the number of data points, R and Rd are the one- dimensional and d-dimensional vector space, respectively, and d is the number of input variables. The value of y (i.e. compressive strength) can be solved as y ¼ w,4ðxÞ þ b (1) where w is a vector of weight coefficient, 4(x) is a nonlinear mapping function, and b is a bias constant. The parameters w and b can be determined by minimizing 1 2kwk2 under the following premise (Samui 2008): Fig. 1. Correlations between input parameters based on Pearson correlation co- efficients. Values 1 and -1 mean strong positive and negative correlations, respectively. M. Wei et al. / Journal of Rock Mechanics and Geotechnical Engineering 14 (2022) 1356e1365 1357\n[page 3] jyi \u0001 ½w , 4ðxiÞ þ b\u0003j \u0004 ε (2) where ε is a free parameter considered as a threshold. Two relaxation variables, i.e. xi and x* i , are introduced to penalize the points outside the ε-insensitive band (Smola and Scholkopf, 2004). The best regression line is defined by minimizing 1 2kwk2 þ c Pd i ¼ 1 \u0002xi þ x* i \u0003, where c is the penalty parameter determining the trade-off between the model complexity and the training error, under the following precondition (Samui 2008): yi \u0001 ½w,4ðxiÞ þ b \u0003 \u0004 ε þ xi ½w,4ðxiÞ þ b \u0003 \u0001 yi \u0004 ε þ x* i ) (3) To facilitate the solution, the Lagrangian multipliers (a* i , ai, a* j and aj) are introduced, and the above quadratic programming problem (Eq. (3)) is transformed into a dual problem of maximizing Pd i ¼ 1yiða* i \u0001aiÞ \u0001 Pd i ¼ 1εða* i þaiÞ \u0001 1 2 Pd i ¼ 1 Pd j ¼ 1ða* i \u0001aiÞða* j \u0001ajÞxT i xj under the following premise: X d i ¼ 1 a* i \u0001 X d i ¼ 1 ai ¼ 0 0 \u0004 ai \u0004 c 0 \u0004 a* i \u0004 c 9 > > > > > = > > > > > ; (4) Fig. 2. A flowchart summarizing how to apply SVM, BPNN, and RF algorithms to establish machine learning models. M. Wei et al. / Journal of Rock Mechanics and Geotechnical Engineering 14 (2022) 1356e1365 1358\n[page 4] By solving Eq. (4), the function model for support vector regression can be determined as y ¼ X d i ¼ 1 \u0002a* i \u0001 ai \u0003Kðxi; xÞ þ b (5) where K (xi,x) is the kernel function, adopting the radial basis function to address nonlinear problems (Samui, 2008): Kðxi; xÞ ¼ eð\u0001gkxi\u0001xk2Þ (6) As summarized in Fig. 2, the two parameters (i.e. c and g) in the radial basis function should be optimized. The SVM-based model aims to find the optimum c and g values through the grid search method (Cheung et al., 1997) and to achieve good generalization capacity and high calculation speed. 2.2.2. BPNN algorithm The BPNN algorithm developed by Rumelhart et al. (1986) is a multi-layer feed-forward algorithm."}], "source_title": "Application of machine learning in predicting the rate-dependent compressive strength of rocks", "source_doi": "10.1016/j.jrmge.2022.01.008", "source_url": "https://www.sciencedirect.com/science/article/pii/S167477552200049X", "source_license": "cc-by-sa", "zotero_item_key": "KNCWQQJD"}
{"id": "656f9c62bd410ccb", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Automated Excavator Based on Reinforcement Learning and Multibody System Dynamics; DOI: 10.1109/access.2020.3040246; license: cc-by-sa."}, {"role": "assistant", "content": "In this study, hydraulic valves are modelled by semi-empirical approach [20]. In this approach, flow through a throttle valve can be described as: Q =    Cv √|dp|, dp > 0 0, dp = 0 −Cv √|dp|, dp < 0 (16) where Cv is the flow rate constant of the valve, and dp is pressure difference between the input and output. Flow through a directional valve can be described as [21]: Q =    CvU√|dp|, dp > 0 0, dp = 0 −CvU√|dp|, dp < 0 (17) where U is the relative spool position, which is described [21]: ˙U = Uref −U τ (18) where Uref is a control signal and τ is a time constant describ- ing the spool dynamics of the valve. D. DEFORMABLE GROUND Together with the multibody approach and hydraulics sim- ulation, the deformable ground forms the basis for devel- opment of the autonomous agents. The deformable ground provides a representation of soil behavior and allows sim- ulation of working scenarios of the machine. In this study, the environment of the excavator is described by employing a combination of the cellular automata-based [22], [23] and particle-based methods, as in [24], [25]. This combination of techniques defines an accurate deformable ground and enables ground material to be moved from one location to another as shown in Fig.2. FIGURE 2. Deformable ground working principle [25]. In the cellular automata-based method [22], [23], the soil is described as cells in a grid (heightfield) which allows the formation of the landscape, as shown in Fig. 2. The cells are vertically divided into piles of blocks. The pressure over the base of one cell is obtained as a combination of the pressure received from their own block and the pressure from a finite number of nearby blocks above the base cell. In this model, the displacement of the soil can occur for two reasons: 1) A large difference in the heightfield of two neigh- boring cells, resulting in avalanches; and 2) A large difference between the vertical forces applied to the neighboring cells, resulting in soil compression and displacement. In this study, the deformation of the heightfield depends on the interaction between the ground and external objects, such as an excavator VOLUME 8, 2020 214001\n[page 5] I. Kurinov et al.: Automated Excavator Based on RL and Multibody System Dynamics bucket or tracks. Depending on the amount of force induced on the soil, the cellular automata redefine the heightfield of the cells, i.e. adding the vertical deformation. The particle-based method allows one to simulate the behavior of the soil when it is sheared with an external force [25]. In this method, particles are generated when the horizontal component of the applied force exceeds the shear impulse limit. During this event, the portion of the cor- responding heightfield is substituted with spherical shaped particles, which have six degrees of freedom. The dynam- ics of the particles can be expressed by methods discussed in the multibody system dynamics subchapter [25]. Particle swarm has a void factor, which is responsible for the space between generated particles, i.e. soil compaction [24]. The soil compaction allows accurate dynamics representation of the soil compared with nature [24]. When the particles reach an equilibrium state, they merge back to the heightfield with a volume update [24]. E. DESCRIPTION OF APLICATION PROGRAMMING INTERFACE RL training and simulation are performed by separate pro- grams. Therefore, a connection between the simulation mod- els and the reinforcement learning algorithm was developed using a custom Application Programming Interface (API). In this study, the API was implemented using the ZeroMQ asynchronous messaging library to send data between the applications. The client-server messaging pattern, which allows multiple clients to be connected to the server [26], was used as the base of the API. The RL side was used as a server and the simulation side as a client. The server side is shown in Algorithm 1 and the simulation side is shown in Algorithm 2. This structure allows simultaneous use of multiple simulations for training as depicted in Fig.3. FIGURE 3. Structure of the API based on Gym, State and Parameters classes. The API consists of the three main classes: Gym, Parame- ters and State. The Gym class is used to control the simulation and handle the connection to the algorithm. The class consists of the make, step and reset methods. The make method is used for initialization of the variables in the simulation."}], "source_title": "Automated Excavator Based on Reinforcement Learning and Multibody System Dynamics", "source_doi": "10.1109/access.2020.3040246", "source_url": "https://ieeexplore.ieee.org/ielx7/6287639/6514899/09268069.pdf", "source_license": "cc-by-sa", "zotero_item_key": "7F2IQNFD"}
{"id": "74f2130515358e0b", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: How to evaluate performance of prediction methods? Measures and their interpretation in variation effect analysis; DOI: 10.1186/1471-2164-13-s4-s2; license: cc-by."}, {"role": "assistant", "content": "Cotton RG, Al Aqeel AI, Al-Mulla F, Carrera P, Claustres M, Ekong R, Hyland VJ, Macrae FA, Marafie MJ, Paalman MH, et al: Capturing all disease-causing mutations for clinical and research use: toward an effortless system for the Human Variome Project. Genet Med 2009, 11:843-849. 26. Kohonen-Corish MR, Al-Aama JY, Auerbach AD, Axton M, Barash CI, Bernstein I, Beroud C, Burn J, Cunningham F, Cutting GR, et al: How to catch all those mutations–the report of the third Human Variome Project Meeting, UNESCO Paris, May 2010. Hum Mutat 2010, 31:1374-1381. 27. Baldi P, Brunak S, Chauvin Y, Andersen CA, Nielsen H: Assessing the accuracy of prediction algorithms for classification: an overview. Bioinformatics 2000, 16:412-424. 28. Fawcett T: An introduction to ROC analysis. Pattern Recognition Letters 2006, 27:861-874. 29. Sonego P, Kocsor A, Pongor S: ROC analysis: applications to the classification of biological sequences and 3D structures. Brief Bioinform 2008, 9:198-209. 30. Giardine B, Riemer C, Hefferon T, Thomas D, Hsu F, Zielenski J, Sang Y, Elnitski L, Cutting G, Trumbower H, et al: PhenCode: connecting ENCODE data with mutations and phenotype. Hum Mutat 2007, 28:554-562. 31. Piirilä H, Väliaho J, Vihinen M: Immunodeficiency mutation databases (IDbases). Hum Mutat 2006, 27:1200-1208. 32. Olatubosun A, et al: PON-P: Integrated predictor for pathogenicity of missense variants. Human Mutation [http://onlinelibrary.wiley.com/doi/ 10.1002/humu.22102/pdf]. 33. Kumar MD, Bava KA, Gromiha MM, Prabakaran P, Kitajima K, Uedaira H, Sarai A: ProTherm and ProNIT: thermodynamic databases for proteins and protein-nucleic acid interactions. Nucleic Acids Res 2006, 34:D204-206. 34. Lonquety M, Lacroix Z, Chomilier J: Pattern recognition in bioinformatics. Heidelberg: Springer; 2008. 35. Tastan O, Yu E, Ganapathiraju M, Aref A, Rader AJ, Klein-Seetharaman J: Comparison of stability predictions and simulated unfolding of rhodopsin structures. Photochem Photobiol 2007, 83:351-362. doi:10.1186/1471-2164-13-S4-S2 Cite this article as: Vihinen: How to evaluate performance of prediction methods? Measures and their interpretation in variation effect analysis. BMC Genomics 2012 13(Suppl 4):S2. Submit your next manuscript to BioMed Central and take full advantage of: • Convenient online submission • Thorough peer review • No space constraints or color figure charges • Immediate publication on acceptance • Inclusion in PubMed, CAS, Scopus and Google Scholar • Research which is freely available for redistribution Submit your manuscript at www.biomedcentral.com/submit Vihinen BMC Genomics 2012, 13(Suppl 4):S2 http://www.biomedcentral.com/1471-2164/13/S4/S2 Page 10 of 10"}], "source_title": "How to evaluate performance of prediction methods? Measures and their interpretation in variation effect analysis", "source_doi": "10.1186/1471-2164-13-s4-s2", "source_url": "https://bmcgenomics.biomedcentral.com/counter/pdf/10.1186/1471-2164-13-S4-S2", "source_license": "cc-by", "zotero_item_key": "LKJJ9WPZ"}
{"id": "02c2500f383b4e37", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Application of machine learning in predicting the rate-dependent compressive strength of rocks; DOI: 10.1016/j.jrmge.2022.01.008; license: cc-by-sa."}, {"role": "assistant", "content": "The process of constructing the BPNN includes signal forward propagation and error back- propagation. The structure of the BPNN consists of an input layer, one or more hidden layers, and an output layer (Fig. 4). In this study, the compressive strength is set as the output neuron, and the others (e.g. strain rate, P-wave velocity) are the inputs. To accelerate the convergence of the neural network, the raw data (X) are normalized using Eq. (7) and mapped into a range of 0e1: Xi ¼ X \u0001 Xmin Xmax \u0001 Xmin (7) where Xi is the normalized data; and Xmax and Xmin are the maximum and minimum values, respectively. The number of hidden-layer nodes has a significant influence on the BPNN training and is determined as h ¼ ffiffiffiffiffiffiffiffiffiffiffi d þ k p þ a (8) where h and k are the number of nodes in the hidden and output layers, respectively, and a is a constant with a value of 0e10. In our model, d ¼ 7 and k ¼ 1. By taking a ¼ 8, h can be determined as 10. Information transfer between the layers relies on the matrix operation: y g ¼ x g \u0005 w g þ b g (9) where x g is the eigenvector of the previous layer, y g is the vector output to the next layer, and w g and b g are the weight and bias matrixes used to convert the eigenvector to the vector output, respectively. In the error back-propagation algorithm, the Sigmoid function is used to introduce nonlinearity to the BPNN-based model to better solve complex nonlinear problems: fsigðzÞ ¼ 1 1þe\u0001z (10) In the initial BPNN-based model, the weights and biases are randomly initialized. After being fed with the input data, the model compares the output data using the loss function: floss ¼ ðby \u0001 yÞ2 (11) where floss is the prediction error, and by and y are the measured and the predicted strength values, respectively. Because of the random parameter setting, a learning mechanism should be implemented to optimize the model. The BPNN-based model has the two parameters (i.e. w g and b g ) to be adjusted during the training process (Fig. 2). Based on the pre- diction error, the error term of each hidden layer is calculated in reverse order through the error back-propagation algorithm, and the gradient descent method is applied to adjust the two param- eters of each hidden layer (Hochreiter et al., 2001). The adjustment of the weight and bias values is determined by the gradient of each parameter calculated using the loss function. The dropout regula- rization method is used to prevent the model from overfitting (Srivastava et al., 2014). An unacceptable error is propagated backwards again, and the weight and bias values are adjusted accordingly. This process is repeated until finding an acceptable error or reaching the maximum cycle number. 2.2.3. RF algorithm The RF algorithm proposed by Breiman (2001) can integrate multiple decision trees into an ensemble algorithm and average the results of these decision trees as the output (Fig. 5). The RF has a fast convergence speed and can prevent data from overfitting to improve the prediction accuracy. The main features of the RF include (i) bootstrap sampling, (ii) random feature selection, (iii) Fig. 3. A schematic diagram of the principle of SVM algorithm. Fig. 4. A schematic diagram of the principle of BPNN algorithm. M. Wei et al. / Journal of Rock Mechanics and Geotechnical Engineering 14 (2022) 1356e1365 1359\n[page 5] out-of-bag (OOB) error estimation, and (iv) full-depth decision tree growth (Li et al., 2018). The training data set (Sm) containing m observations can be written as Sm ¼ fðx1; y1Þ; ðx2; y2Þ; :::; ðxm; ymÞg ðxi ˛ Rd; yi ˛ RÞ (12) where m is the number of training data points. Also shown in Fig. 2, the bootstrap sampling is performed to generate t sets of training data from the original data set. Each new data set corresponds to a decision tree that is fully grown and not pruned by the classification and regression trees (CART) method- ology (Duda et al., 2000). The RF can find the best split among the randomly selected features at each node. It is common to artificially control the number of branches in the decision tree or let it grow freely for a limited number of times and subsequently to seek the best single decision tree. After the training process, a global pre- diction function y ¼ hðx; SmÞ is constructed over Sm. The RF pro- duces the prediction result by averaging the output of these decision trees (i.e. ~y1 ¼ ~hðx; ~S1Þ, ~y2 ¼ ~hðx; ~S2Þ, ., ~yt ¼ ~hðx; ~StÞ)."}], "source_title": "Application of machine learning in predicting the rate-dependent compressive strength of rocks", "source_doi": "10.1016/j.jrmge.2022.01.008", "source_url": "https://www.sciencedirect.com/science/article/pii/S167477552200049X", "source_license": "cc-by-sa", "zotero_item_key": "KNCWQQJD"}
{"id": "14cd567df2b10629", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Two-phase multi-model automatic brain tumour diagnosis system from magnetic resonance images using convolutional neural networks; DOI: 10.1186/s13640-018-0332-4; license: cc-by-sa."}, {"role": "assistant", "content": "Further improve- ments have been realized to classify an abnormal image as benign (noncancerous) or malignant (cancerous) by applying a double classifier Abd-Ellah et al. [19]. Devasena and Hemalatha [20] proposed a CAD system for the detec- tion of abnormal parts in MRIs using a hybrid abnormality detection algorithm (HADA). In [21], Patil and Udupi used preprocessing, segmenta- tion, feature extraction, and a probabilistic neural network (PNN) method to identify brain tumours. Arakeri and Reddy [22] applied three different classifiers: an SVM, an artificial neural network (ANN), and a k-nearest neigh- bour (k-NN) algorithm to distinguish between different types of tumours. Dandıl et al. [23] proposed a system for brain tumour classification where spatial-fuzzy C-means (FCM) was applied for brain tumour segmentation, and an SVM was used in the classification stage. Goswami and Bhaiya [24] presented a new MRI brain tumour classification based on an ANN. Histogram equalization, noise filtering, and edge detection were applied in the preprocessing stage. The features were extracted by independent component analysis (ICA). A self-organizing map neural network (SOMNN) was\n[page 3] Abd-Ellah et al. EURASIP Journal on Image and Video Processing (2018) 2018:97 Page 3 of 10 applied in the classification stage. Deepa and Devi [25] proposed a system consisting of feature extraction, clas- sification, and tumour segmentation. Optimal texture features were extracted from tested images using statis- tical features. A back propagation neural network (BPN) and a radial basis function neural network (RBFN) were used in the classification and segmentation stages, respec- tively. A short comparison of the aforementioned classical learning-based diagnosis methods is provided in Table 1. Few deep learning-based methods for brain tumour diagnosis were found in the state-of-the-art in this field. In [26], brain CT image classification based on a deep neural network (DNN) was presented. A gray-level co- occurrence matrix (GLCM) was used for feature extrac- tion, and a DNN was applied for classification. The aver- age accuracy of the reported method was 83%. Gao et al. [27] introduced a CT brain image classification that inte- grated both 2D and 3D CNNs. The average classification accuracy rate was 87.6%, while each network consisted of seven layers. Yan Xu et al. proposed a method using deep convolutional activation features for brain tumour classification and segmentation. The system classification accuracy was 97.5% [28]. Several deep learning methods for brain tumour local- ization have been presented in the literature. Havaei et al. [29] demonstrated automatic brain tumour localization using a DNN based on the CNNs with an average DICE score of 0.88. In [30], a fully convolutional residual neural network (FCR-NN) combined with a fully convolutional architec- ture with optimization gains from residual identity was developed. The network consists of a 22 complex layers, and the achieved DICE score was 0.87. Casamitjana et al. Table 1 Comparison of the available traditional classification approaches used in brain tumour detection CAD system Total images Classification method Accuracy (%) El-Dahshan et al. [2] 101 FFBPNN 99.00 Devasena and Hemalatha [20] 250 HADA 98.8 Arakeri and Reddy [22] 550 SVM, ANN and k-NN 99.09 Dandıl et al. [23] 376 KSVM 91.49 Goswami and Bhaiya [24] 70 SOMNN 98.60 Deepa and Devi [25] 42 BPN and RBFN 85.71 Abd-Ellah et al. [18] 80 KSVM 100 Abd-Ellah et al. [19] 120 KSVM 100 [31] proposed the use of 3D CNNs for the segmenta- tion of abnormal parts in MRIs based on a network using multi-resolution features and reported a DICE score of 0.84. Zhao et al. [32] used a fully convolutional neural net- work (FCNN) and conditional random fields (CRFs) as a post-processing step to segment brain tumours with an obtained DICE score of 0.87. Pereira et al. in [33] presented a preliminary study on brain tumour segmen- tation based on fully convolutional network (FCN) with standardization of MRI histograms as a preprocessing stage, which demonstrated a DICE score of 0.85. More recently, Pereira et al. [34] examined a complete seg- mentation system that includes preprocessing, CNN, and post-processing stages. The new system could achieve an improved DICE score of 0.88 at the expense of increased system complexity. 3 Methods Currently, deep learning models are of considerable research importance."}], "source_title": "Two-phase multi-model automatic brain tumour diagnosis system from magnetic resonance images using convolutional neural networks", "source_doi": "10.1186/s13640-018-0332-4", "source_url": "https://jivp-eurasipjournals.springeropen.com/track/pdf/10.1186/s13640-018-0332-4", "source_license": "cc-by-sa", "zotero_item_key": "2ZZSU54R"}
{"id": "a8d7cccd372f4a99", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Application of machine learning in predicting the rate-dependent compressive strength of rocks; DOI: 10.1016/j.jrmge.2022.01.008; license: cc-by-sa."}, {"role": "assistant", "content": "We then established three machine learning models based on SVM, the BPNN, and the RF algorithms. In the machine learning models, the compressive strength was the output, and the other properties (i.e. static compressive strength, P- wave velocity, specimen dimension, grain size, bulk density, and strain rate) were the inputs. After training and validating the models, we discussed the importance of each input parameter in the strength prediction and the transitional behaviors of rocks with different specimen dimensions. Our results indicated that the machine learning models can achieve reliable prediction for the rate-dependent strength of rocks. The RF algorithm manifests slight superiority to SVM and BPNN in predicting the rock strength. The strain rate as input is more important than the other parameters to ensure the excel- lent performance of the models. The dynamic increase factor is independent of specimen dimension at low strain rates but bifurcate for the strain rate over a relatively high level. This observation can be used to study the transitional behaviors of rocks from low to high strain rates, which may yield more in- sights into the specimen dimension effect on rock strength and further highlight the strain rate effect. This study introduced a new avenue to study the complex mechanical behaviors of rocks at various strain rates. Fig. 9. Effect of strain rate on (a) Compressive strength, and (b) Dynamic increase factor for different specimen diameters (12.5mm, 25mm, and 50 mm), predicted by the RF-based model. Fig. 10. Effect of strain rate on (a) Compressive strength, and (b) Dynamic increase factor for different aspect ratios (0.5, 0.75, and 1), predicted by the RF-based model. M. Wei et al. / Journal of Rock Mechanics and Geotechnical Engineering 14 (2022) 1356e1365 1364\n[page 10] Declaration of competing interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Acknowledgments This research is supported by National Research Foundation, Singapore under its Virtual Singapore R&D Programme (Award No. NRF2019VSG-GMS-001). Appendix A. Supplementary data Supplementary data to this article can be found online at https://doi.org/10.1016/j.jrmge.2022.01.008. References Boser, B.E., Guyon, I.M., Vapnik, V.N., 1992. A training algorithm for optimal margin classifiers. In: Proceedings of the Fifth Annual Workshop on Computational Learning Theory, pp. 144e152. Pittsburgh Pennsylvania, USA. Breiman, L., 2001. Random forests. Mach. Learn. 45 (1), 5e32. Chen, W.W., Song, B., 2011. Split Hopkinson (Kolsky) Bar: Design, Testing and Ap- plications. Springer, New York. Cheung, B.K.S., Langevin, A., Delmaire, H., 1997. Coupling genetic algorithm with a grid search method to solve mixed integer nonlinear programming problems. Comput. Math. Appl. 34 (12), 13e23. Drucker, H., Burges, C.C., Kaufman, L., Smola, A.J., Vapnik, V.N., 1996. Support vector regression machines. Adv. Neural Inf. Process. Syst. 28 (7), 779e784. Du, H.B., Dai, F., Xu, Y., Liu, Y., Xu, H.N., 2018. Numerical investigation on the dy- namic strength and failure behavior of rocks under hydrostatic confinement in SHPB testing. Int. J. Rock Mech. Min. Sci. 108, 43e57. Duan, K., Ji, Y.L., Wu, W., Kwok, C.Y., 2019a. Unloading-induced failure of brittle rocks and implications for excavation-induced strain burst. Tunn. Undergr. Space Technol. 84, 495e506. Duan, K., Li, Y.C., Wang, L., Zhao, G.F., Wu, W., 2019b. Dynamic responses and failure modes of stratified sedimentary rocks. Int. J. Rock Mech. Min. Sci. 122, 104060. Duda, R.O., Hart, P.E., Stork, D.G., 2000. In: Pattern Classification, third ed. John Wiley & Sons, New York, USA. Edwards, A.L., 1976. An Introduction to Linear Regression and Correlation. W. H. Freeman. Fan, L.F., Gao, J.W., Wu, Z.J., Yang, S.Q., Ma, G.W., 2018. An investigation of thermal effects on micro-properties of granite by X-ray CT technique. Appl. Therm. Eng. 140, 505e519. Fan, L.F., Gao, J.W., Du, X.L., Wu, Z.J., 2020. Spatial gradient distributions of thermal shock-induced damage to granite. J. Rock Mech. Geotech. Eng. 12 (5), 917e926. Frew, D.J., Forrestal, M.J., Chen, W., 2001. A split Hopkinson pressure bar technique to determine compressive stress-strain data for rock materials. Exp. Mech. 41 (1), 40e46. Gamal, H., Alsaihati, A., Elkatatny, S., Haidary, S., Abdulraheem, A., 2021."}], "source_title": "Application of machine learning in predicting the rate-dependent compressive strength of rocks", "source_doi": "10.1016/j.jrmge.2022.01.008", "source_url": "https://www.sciencedirect.com/science/article/pii/S167477552200049X", "source_license": "cc-by-sa", "zotero_item_key": "KNCWQQJD"}
{"id": "b04420cf023ed16b", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Detection of Ionospheric Scintillation Based on XGBoost Model Improved by SMOTE-ENN Technique; DOI: 10.3390/rs13132577; license: cc-by-sa."}, {"role": "assistant", "content": "2021, 13, 2577. https://doi.org/10.3390/rs13132577 https://www.mdpi.com/journal/remotesensing\n[page 2] Remote Sens. 2021, 13, 2577 2 of 22 be related to the magnetic local time, resulting in the fast movement of large-scale iono- spheric plasma structures and phase fluctuations [8–10]. In low-latitude areas, equatorial scintillation activities are stronger and frequent during years with solar maxima, which are associated with the solar activities [11]. The GNSS receiver is affected by amplitude and phase fluctuation, carrier Doppler jitter, cycle slips, and the serious loss of lock and even navigation interruption, which have adverse effects on both ionospheric research and navigation and positioning services [12,13]. Evidence suggests that ionospheric scintillation plays a vital role in receiver perfor- mance in positioning and navigation services. To improve the quality and reliability of GNSS observations, the implementation of the automatic detection and monitoring of ionospheric scintillation is vital, especially during strong scintillation. Relevant ionospheric scintillation monitoring receivers (ISMRs) have recently been designed for use in commerce and research, which are used by presetting thresholds for amplitude and phase indices [14]. However, detection methods such as traditional thresholding technologies are influenced by various factors. To be specific, this approach requires detrending and filtering algo- rithms for observations and overlooks higher-order moment information of the GNSS signals [15]. Moreover, it is sensitive to false alarms due to factors such as multipath, and has been proved to present a lower detection accuracy of 81%, compared to the approach of manual visual inspection [15,16]. Some wavelet decomposition and transform-based techniques using Butterworth filters with non-indices are proposed as alternatives to over- come the problem of detrending, but they rely on expensive computation and complex implementation [17,18]. In recent years, researchers have investigated a variety of machine learning approaches to achieve automatic scintillation detection. Jiao et al. (2017) [19,20] firstly exploited the support vector machine (SVM) for amplitude scintillation detection on two classes. They used a mass of real data and manual labels in the training process and achieved a detection accuracy of 91–96%, outperforming other traditional triggering approaches and non-index techniques. A similar detection method for phase scintillation presents an accuracy of around 92% [21,22]. Lin et al. (2020) [23] analyzed the effects of a binary classification SVM model on hyperparameters and achieved excellent performance in testing. However, based on features such as scintillation indices and relevant maximum and average values, these methods require a filtering process for the scintillation indices S4/SigmaPhi and predeter- mined elevation mask of 30◦. Ludwig-Barbosa, V. et al. (2021) [24] trained SVM models with features combining amplitude/phase scintillation indices along with corresponding maximum and mean values, as well as intensity power spectral density (PSD), showing about 91% accuracy in the detection of ionospheric scintillation. Besides the detection of ionospheric scintillation by an SVM, similar research such as that on jamming detection in GNSS bands with an SVM was conducted, with 94.4% accuracy [25]. An accuracy of 91.36% was found when performing a similar task using convolutional neural networks [26], which are widely researched for various classification tasks, showing great performance [26–28]. Linty et al. (2019) [16] proposed a decision tree algorithm relying on the in-phase and quadrature correlator outputs of the receiver tracking loop, which are considered as sample features after simple computation. The results of 10-fold cross-validation show that the accuracy of amplitude scintillation detection reached 96.7% for features consisting of S4, the carrier-to-noise rate (C/N0) and satellite elevation. Furthermore, for features used by correlator outputs and corresponding combinations, the cross-validation accuracy in- creased from 98.0% to 99.7% when using a random forest algorithm. The overall F1 score reached a high value of 90%, compared with the value of 80.1% achieved with the semi- hard rule [16]."}], "source_title": "Detection of Ionospheric Scintillation Based on XGBoost Model Improved by SMOTE-ENN Technique", "source_doi": "10.3390/rs13132577", "source_url": "https://www.mdpi.com/2072-4292/13/13/2577/pdf?version=1625212202", "source_license": "cc-by-sa", "zotero_item_key": "4YEIWMK3"}
{"id": "eec5dc1815dc95a4", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Detection of Ionospheric Scintillation Based on XGBoost Model Improved by SMOTE-ENN Technique; DOI: 10.3390/rs13132577; license: cc-by-sa."}, {"role": "assistant", "content": "General Overviews This section presents brief descriptions of GNSS ionospheric scintillation, the data collection system and the feature extraction process.\n[page 4] Remote Sens. 2021, 13, 2577 4 of 22 2.1. GNSS Ionospheric Scintillation GNSS satellites transmit low-rate navigation messages modulated by unique spread- ing sequences or codes, which are carried by radio frequency (RF) signals and received by GNSS receivers on the Earth. Corresponding computations on position, velocity and time can be utilized in land applications such as autonomous vehicle navigation and vehicle tracking monitoring, in marine applications such as ocean transportation and inland river shipping, and in aviation applications such as route navigation, airport scene monitoring and precision approach. In the propagation of RF signals from satellites to receivers, there are natural distur- bances in the GNSS signals when they pass through the ionospheric irregularity, resulting in temporal delay and fluctuations, degrading the accuracy, integrity and reliability of the system’s performance. Figure 1 shows the effects of ionospheric scintillation on the process of transiting satellite signals to GNSS receivers. Ionospheric scintillation, a form of major interference, is difficult to model due to its quasi-random nature. With the increased de- mand for and dependence on navigation systems in various fields, research on improving the accuracy of scintillation detection is significant. Ionospheric irregularity Rotaion polarization Multipath effect Figure 1. Representative effects of ionospheric scintillation on GNSS receiver. The green blocks refer to the ionospheric irregularities on the area of ionospheric layer, and the red lines represent the main disturbances on the ground receiver. Ionospheric scintillation is classified into amplitude scintillation and phase scintilla- tion; the former is more frequent in low-latitude areas. It is extremely important for a better detection method to be developed for the Beidou navigation system (BDS), which has been built in recent years with an immature scintillation detection technique [31]. Based on existing GPS data consisting of strong scintillation that were collected in the latest peak year for solar activity, relevant research on scintillation detection can be carried out to develop corresponding technology for the BDS. The performance can also be evaluated in the coming peak year for solar activity (2023–2024). S4, commonly used as an indicator, ranges from 0 to 1, for which a large value represents a stronger intensity of ionospheric scintillation. Values greater than 1 represent the occurrence of extremely severe ionospheric scintillation. The scintillation levels are normally classified into three categories [32–34]: strong if 0.6 ≤S4, moderate if 0.2 < S4 ≤0.6, and weak if 0 < S4 ≤0.2. Due to the occasional contingency, irregularity and certain seasonal characteristics of ionospheric scin- tillation, there are far fewer strong scintillation events than weak and moderate scintillation events [35]. Moreover, most of the scintillation activity over the course of a day occurs from sunset to dawn [36]. In other words, the proportion of strong scintillation events in with respect to scintillation events of all intensities is relatively low, causing imbalance [16,20,23].\n[page 5] Remote Sens. 2021, 13, 2577 5 of 22 Thus, the imbalance exists naturally in ionospheric scintillation detection and monitoring. However, such a phenomenon has not been taken into consideration for the detection of interference in space weather monitoring. 2.2. Data Collection System The data used in this paper were collected at the low-latitude site of São José dos Cam- pos, Brazil (23.2S, 45.9W), from 2013 to 2015 during the last peak of the solar cycle, where the phenomenon of ionospheric scintillation was extremely active due to its geographical location close to both the South Atlantic Magnetic Anomaly (SAMA) and the EIA [37]. Months of GPS data from 2013 to 2015 were recorded, with strong ionospheric scintillation detected during each hour of monitoring. Aiming at studying the natural, occurrent and unpredictable phenomenon, a Septentrio PolaRx ionospheric scintillation monitoring (ISM) receiver was utilized to monitor the scintillation activity. Figure 2 shows the framework of the data collection system used to monitor ionospheric scintillation events."}], "source_title": "Detection of Ionospheric Scintillation Based on XGBoost Model Improved by SMOTE-ENN Technique", "source_doi": "10.3390/rs13132577", "source_url": "https://www.mdpi.com/2072-4292/13/13/2577/pdf?version=1625212202", "source_license": "cc-by-sa", "zotero_item_key": "4YEIWMK3"}
{"id": "fce0dd31d9b52f04", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Spectral Angle Mapping and AI Methods Applied in Automatic Identification of Placer Deposit Magnetite Using Multispectral Camera Mounted on UAV; DOI: 10.3390/min12020268; license: cc-by-sa."}, {"role": "assistant", "content": "In principle, this meant that more multispectral images were to be captured at 2 m with a smaller spatial area per image, followed by 10 m, and the least number of images being captured at 20 m with a relatively larger spatial area. The drone was flown over the 30 m × 90 m area in order to pre-program the flight plan. This flight setting procedure ensures the drone flies within the desired area, direction and elevation before allowing it to perform all these maneuvers autonomously without human intervention. 2.2. Parallax Error Correction Since there is a small distance between the 6 spectral sensors of the drone, parallax error always needs to be corrected for. This ensures that all sensors are aimed at the same image scene depending on the UAV flight height. At low altitudes, the parallax error is significantly higher relative to higher flight altitudes [23]. There are several methods\n[page 5] Minerals 2022, 12, 268 5 of 19 such as Hough transform and least-squares techniques amongst others which can be applied to perform this task aimed at interpolating the images into perfectly overlapping images [23,24]. However, according to Laurence [24], the edge detection method performs well as it takes into account the individual pixels within each of the 6 multispectral images to be overlapped. It calculates the centers of gravity, and automatically nullifies the differences in position, thereby creating a perfect overlay of images, hence the application of this method in this study [24]. It should be noted however, that in order to assume a perfect overlay, each of the 6 images loses a small part of their edges. The lower the UAV elevation, the more data are lost relative to higher flight elevations. The data loss depends on the distance from the drone sensor to the subject, as well as the distance between sensors within the drone, and so, parallax error always needs to be corrected. Though a small percentage of data is lost per image during parallax correction, this is not the case when considering the study area as a whole. The reason for this is that the UAV flight missions were set up such that each captured image slightly overlaps those around it. This, in principle, ensures zero data loss within the 30 m × 90 m study area, with only the image regions outside the study area experiencing data loss. 2.3. A Scrutiny of SAM Analysis Based on a number of vegetation criteria, Mohajane et al. [4] have employed several indexes in an attempt to perform rapid field vegetation classification. These included the intensity of certain vegetation colours and moisture content via SAM. Zhang and Li [8] have used the same principle in employing SAM as an index to classify lithology. Weyermann et al. [11] define SAM as a physically based spectral classification that uses an n-dimensional angle to match pixels to reference spectra (Figure 3). The algorithm determines the spectral similarity between spectra by deducing the angle between them, treating them as vectors in a space with dimensionality equal to the number of bands [17], which in this study is 6 bands. This technique, when used on calibrated reflectance data as done in this study, is relatively insensitive to illumination and albedo effects [16]. For this reason, SAM is highly applicable in multitudes of real life field observations, hence the motivation to employ this technique in this study. In short, SAM compares the angle between the endmember spectrum vector and each pixel vector in the n-dimensional space. Narrow angles represent closer matches to the reference spectrum [8]. Pixel data further away from the specified maximum angle threshold in radians are not classified as the reference spectrum of what is sought after, which in this study is magnetite iron sand. Figure 3 illustrates the principle of SAM analysis via cosine similarity. Figure 3. Schematic diagram showing the principle of spectral angle mapping via cosine similarity. Reference spectrum is the sought after target. Angle a is the cosine similarity threshold between a reference and a true spectrum.\n[page 6] Minerals 2022, 12, 268 6 of 19 The drone is linked to a tablet held by the operator on the ground, it receives captured data allowing for on-site SAM analysis, aiding in the identification of magnetite iron sands. Though best performed in the field, this operation can be performed in laboratory analytics. To perform SAM, each captured multispectral image is imported into a SAM operation algorithm such as python."}], "source_title": "Spectral Angle Mapping and AI Methods Applied in Automatic Identification of Placer Deposit Magnetite Using Multispectral Camera Mounted on UAV", "source_doi": "10.3390/min12020268", "source_url": "https://www.mdpi.com/2075-163X/12/2/268/pdf?version=1645347690", "source_license": "cc-by-sa", "zotero_item_key": "VPIE7856"}
{"id": "e6fe7f8ad597c87f", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Reinforcement learning based process optimization and strategy development in conventional tunneling; DOI: 10.1016/j.autcon.2021.103701; license: cc-by."}, {"role": "assistant", "content": "Deep Q-learning is a deep reinforcement learning technique that extends the capabilities of clas­ sical Q-learning [34,35] by replacing the value iteration in the Q-table with the function approximator of deep artificial neural networks (ANN). Although applications of ANNs for geotechnical purposes are still often seen as complementary to conventional computational models [36,37], in this case the use of ANNs allows for applications in complex and continuous states spaces while classical Q-learning is confined to discrete states. DQN algorithms are off-policy, model free RL techniques following the Bellman equation (Eq. (1) after [14]) where the optimal action-value function Q*(s,a) is based on a state s and after having taken an action a. The best action is chosen by maximizing the expected value of r + γQ*(s′,a′) where r is the reward, γ is the discount factor that determines how important the future reward is to the algorithm (see also Table 2), s′ is the state at the next time step and a′ are all possible actions [14]. Q*(s, a) = Es′ [ r + γmax a′ Q*(s ′, a ′) |s, a ] (1) [14] have shown that the DQN algorithm can be used for a wide range of different RL problems and albeit there are improvements to deep Q- learning we chose the original implementation as we see it as well suited to establish a baseline for further developments in geotechnical RL. Our implementation is based on the DQN after [14] and the custom DQN implementation from [38]. The main deviations of the network archi­ tecture in comparison to [14,38] are due to the shape of the input and output data with the input being an 2 × 2100 × 2 array and the output a vector of length 8. The number of hidden layers and the decreasing kernel size from the top to the bottom convolutional layers is in accor­ dance with [14] as is the size of the kernel’s stride which is half the size of the kernel itself. While [14] used 32, 64 and 64 filters for each of the three hidden layers respectively, we used 32, 64 and 32 filters for these layers, as we observed that the agent’s performance did not suffer from this decrease, while the computational speed increased. In accordance with these authors we used rectified linear units (ReLU) [39] activation functions. ReLU activation functions have been widely adopted for ANNs within the past decade as they have shown to achieve a better performance than previously used activation functions like the sigmoid (see [40]). As given in [14] the general DQN’s ANN architecture is that of a deep convolutional neural network [41] whose hierarchical struc­ ture mimics the effect of receptive fields and is inspired by [42]. Like [14] we did not perform systematic hyperparameter tuning by random search or similar techniques (see e.g. [43]) due to the big computational effort of the simulation. Hyperparameters were thus optimized manually throughout the course of the development of the RL-simulation. We nevertheless point out that the given DQN architecture as well as the used hyperparameters still have room for improvement (see outlook in Section 7). Table 2 lists all hyperparameters for our DQN implementation. As we use the same terminology as [14] the reader is referred to this paper for more information on the individual parameters. We implemented the DQN using the tensorflow [44] based Python library Keras [45]. Training was done on a NVIDIA GeForce RTX 2080 Ti. From input to output, the agent’s architecture goes as follows and a graphical representation is given in Fig. 5 (with adaptions of architec­ ture from [14] as described in the section ahead): ▪ The input consists of an array with the shape 2 × 2100 × 2 (see Section 3.3.1). ▪ One convolutional layer, with 32 filters, a kernel size of 1 × 16 and a stride of 1 and 8 applying a ReLU activation function [39] ▪ One convolutional layer, with 64 filters, a kernel size of 1 × 8 and a stride of 1 and 4 applying a ReLU activation function ▪ One convolutional layer, with 32 filters, a kernel size of 1 × 4 and a stride of 1 and 2 applying a ReLU activation function ▪ One fully connected layer with 256 neurons applying a ReLU activation function ▪ One fully connected layer with 8 neurons (one per action) as the output layer which applies a linear activation Table 1 The mechanical parameters and permeability of the two ground types, where gt1 represents weak rock and gt2 stronger rock."}], "source_title": "Reinforcement learning based process optimization and strategy development in conventional tunneling", "source_doi": "10.1016/j.autcon.2021.103701", "source_url": "https://www.sciencedirect.com/science/article/pii/S0926580521001527", "source_license": "cc-by", "zotero_item_key": "RXHUVV3X"}
{"id": "f733686311e23184", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: How to evaluate performance of prediction methods? Measures and their interpretation in variation effect analysis; DOI: 10.1186/1471-2164-13-s4-s2; license: cc-by."}, {"role": "assistant", "content": "In the column wise parameters, which are for the ratios of either positive or negative cases (sensitivity and specificity), are not affected whereas there is a significant difference in NPV and PPV, which are row wise ratios based on numbers of both posi- tive and negative cases. In all the examples, 75 % of both positive and negative cases are correctly predicted and therefore sensitivity and specificity remain the same. It is thus apparent that imbalance in class sizes grossly affects the NPV and PPV evaluation criteria. To overcome the class imbalance problem different approaches can be taken. One is to prune the size of the bigger class to be that of the smaller one. It is also possible to normalize in the contingency table the values of either positive or negative cases to have the total of the other class. Quite often in bioinformatics limited amount of data are available and therefore one would be reluctant to delete part of the datasets. When normalizing the data be sure that the existing dataset is representative otherwise bias in the set may further be increased. Accuracy and MCC Specificity, sensitivity, PPV and NPV are calculated by using only half of the information in the contingency table and thus cannot represent all aspects of the performance. Accuracy (Fig. 4) and Matthews correlation coefficient (MMC) take benefit of all the four numbers and as such are more balanced, representative and comprehensive than the line or column wise measures. The MCC is calculated as follows: MCC TP TN FP FN TP FN TN FP TP FP TN FN         ( )( )( )( ) . For all the measures discussed in here applies that higher the value the better. Except for MCC, the values range from 0 to 1. MCC ranges from -1 to 1. -1 indi- cates perfect negative correlation, 0 random distribution and 1 perfect correlation. Accuracy and MCC are affected by class imbalance only in extreme cases. The effect of the correctly predicted cases on the para- meters in equally distributed dataset is shown in Fig. 6. The value for MCC grows slower than the others reach- ing 0.5 when 75 % of cases are correctly predicted. Ran- dom results (50 % of both negative and positive correctly predicted) gives a value of 0, while the other parameters - sensitivity, specificity, PPV, NPV, and accuracy are 0.5. Fig. 6. can be used to check the performance of equally Table 1 Evaluation measures for test data -50 % -25 % Equal +25 % +50 % tp 750 750 750 750 750 fn 250 250 250 250 250 tn 375 563 750 938 1125 fp 125 187 250 312 375 sensitivity 0.75 0.75 0.75 0.75 0.75 specificity 0.75 0.75 0.75 0.75 0.75 PPV 0.86 0.80 0.75 0.71 0.67 NPV 0.60 0.69 0.75 0.79 0.82 accuracy 0.75 0.75 0.75 0.75 0.75 MCC 0.48 0.50 0.50 0.50 0.49 Example of a situation when both positive and negative cases have the same rate of correct predictions (75 %) and when the amount of negative cases is either equal or 25 or 50 % lower or higher than that for positive cases. Figure 6 The growth of the performance measures along increasing reliability Graphs for quality measures for equally distributed data (same amount positive and negative cases) when the performance increases equally in both classes. The solid curve indicates the growth of sensitivity, specificity, PPV, NPV, and accuracy. The dotted line is for MCC. Vihinen BMC Genomics 2012, 13(Suppl 4):S2 http://www.biomedcentral.com/1471-2164/13/S4/S2 Page 6 of 10\n[page 7] distributed dataset if e.g. some parameters in an article are not provided. Biases can easily be seen as deviations from the relationships in the figure. To obtain full picture of the predictor performance it is important to evaluate all the six measures together. Other parameters Several other parameters can be derived from the contin- gency matrix. These are not discussed further as they are not widely used in literature and can be easily calculated from the six previously presented parameters. These include false positive rate (FPR) which equals 1-specificity and false negative rate (FNR) which is 1-sensitivity. False discovery rate (FDR) is 1-PPV. Positive and negative likelihood ratios are calculated as follows: LR LR      sensitivity specificity sensitivity FPR and sensi 1 1 tivity specificity FNR specificity  . F measure is another one that uses all the data. It is calculated as: F     2 PPV sensitivity PPV sensitivity . Other measures include e.g."}], "source_title": "How to evaluate performance of prediction methods? Measures and their interpretation in variation effect analysis", "source_doi": "10.1186/1471-2164-13-s4-s2", "source_url": "https://bmcgenomics.biomedcentral.com/counter/pdf/10.1186/1471-2164-13-S4-S2", "source_license": "cc-by", "zotero_item_key": "LKJJ9WPZ"}
{"id": "4885d3bfde8cf1fc", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Application of machine learning in predicting the rate-dependent compressive strength of rocks; DOI: 10.1016/j.jrmge.2022.01.008; license: cc-by-sa."}, {"role": "assistant", "content": "However, when the strain rate ex- ceeds 50 s\u00011, the DIF value of the specimen with 50 mm diameter increases faster than those of the other specimens. The effect of specimen diameter can be ascribed to an amplified inertia effect (Chen and Song, 2011). The results additionally reveal that the compressive strength and the DIF value exhibit different re- sponses to an increase in strain rate. The difference among the three curves for compressive strength decreases with a higher strain rate, while the three curves for the DIF values bifurcate when the strain rate is greater than 50 s\u00011. This is because the static compressive strength decreases with a larger specimen diameter (Masoumi et al., 2015). Fig. 7. Percentage metrics, including (a) MAE, (b) MXAE, (c) MAPE, and (d) RMSE, to evaluate the performance of the SVM-based, BPNN-based, and RF-based models, respectively. M. Wei et al. / Journal of Rock Mechanics and Geotechnical Engineering 14 (2022) 1356e1365 1362\n[page 8] The RF-based model also considers similar fine-grained sandstone specimens with a fixed diameter (50 mm) and three different aspect ratios (0.5, 0.75, and 1). The results exhibit that the compressive strength slightly decreases with increasing aspect ratio at a given strain rate, probably due to more defects in a longer specimen and weaker end friction effect(Fig.10a). The negative correlation between the compressive strength and specimen length is also observed in other rocks and solid materials (like concrete) (Krauthammer et al., 2003; Munoz and Taheri, 2017). Fig. 10b shows that the aspect ratio has a minor effect on the DIF value for a strain rate lower than 50 s\u00011, but obviously influences the DIF value beyond the strain rate. The fluctuation of thecurvesislikely due tothediscrepancies of input data points, which can be related to the system errors from the test setup and the data acquisition unit. Combining Figs. 9b and 10b, it suggests that the influence of specimen dimension on the strength enhance- ment can beoverlookedindetermination of thecompressive strength at low strain rates (e.g. < 50 s\u00011), but should be examined at inter- mediate and high strain rates. Both the figures also illustrate that the DIF values for different specimen dimensions increase at the same increment rate at low strain rates. However, the increment rate be- comes different at higher strain rates. This observation provides new insights into the transitional behaviors of rocks from the low to high strain rates. The strength enhancement at the beginning of the tran- sitional process is less dependent of the specimen dimension, and the dependence of the specimen dimension becomes obvious with increasing strain rate. Fig. 8. Percentage increments of performance metrics: (a) MAE, (b) MXAE, (c) MAPE, and (d) RMSE, obtained from excluding each input parameters in the SVM- (red solid line), BPNN- (black dotted line), and RF-based (blue dashed line) models, respectively. M. Wei et al. / Journal of Rock Mechanics and Geotechnical Engineering 14 (2022) 1356e1365 1363\n[page 9] Our study highlights that the correlations among various geometrical, physical, and mechanical properties of rocks are nonlinear and very different from the simple correlations based on the Pearson correlation coefficient (Fig. 1). Additionally, the nonlinear correlations can be even more complex if the three models can be trained by the data points obtained from laboratory and field experiments on rocks with larger sizes under higher strain rates. In other words, this study is limited with the data points in small ranges of specimen dimensions and strain rates available in the literature. The established models are applicable to forecast the compressive strength of hard rocks at a strain rate ranging from 5 \u0005 10\u00016 s\u00011 to 223 s\u00011 (Table A1). For other strain rates, additional training with the corresponding data is needed to extend the applicability of these models. The study demonstrates that machine learning models could be an effective method to explore the complexity of rock dynamics, particularly those difficult to control in laboratory experiments. 5. Conclusions This study addressed the feasibility of applying machine learning models to predict the rate-dependent strength of rocks. We first collected the strength data of various types of hard rocks as well as the corresponding geometrical, physical, and mechanical properties in the literature."}], "source_title": "Application of machine learning in predicting the rate-dependent compressive strength of rocks", "source_doi": "10.1016/j.jrmge.2022.01.008", "source_url": "https://www.sciencedirect.com/science/article/pii/S167477552200049X", "source_license": "cc-by-sa", "zotero_item_key": "KNCWQQJD"}
{"id": "6618e437ae6bd037", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Reinforcement learning based process optimization and strategy development in conventional tunneling; DOI: 10.1016/j.autcon.2021.103701; license: cc-by."}, {"role": "assistant", "content": "Lanham, Reinforcement Learning, from Games to Geologic Interpretation. htt ps://medium.com/@cxbxmxcx/reinforcement-learning-from-games-to-geologi c-interpretation-93757664f0e4, 2019 (accessed 2 October 2020). [18] M. Lanham, Reinforcement Learning, It’s Coming not Just for Games. https://me dium.com/@cxbxmxcx/reinforcement-learning-its-coming-not-just-for-games-6b6 064e02bbe, 2019. [19] D. Schwung, A. Schwung, S. Ding, Actor-critic reinforcement learning for energy optimization in hybrid production environments, Int. J. Comput. 18 (2019) 360–371. [20] I. Carlucho, M. de Paula, S. Wang, Y. Petillot, G.G. Acosta, Adaptive low-level control of autonomous underwater vehicles using deep reinforcement learning, Robot. Auton. Syst. 107 (2018) 71–86, https://doi.org/10.1016/j. robot.2018.05.016. [21] B. Maidl, M. Thewes, U. Maidl, D. Sturge, Handbook of Tunnel Engineering, First, Engl. ed., Ernst/Wiley, Berlin, 2013. [22] W. Stipek, R. Galler, M. Bauer (Eds.), 50 Years of NATM: Experience Reports, ITA, Austria, Wien, 2012. [23] Norwegian Tunnelling Society (Ed.), Norwegian Tunnelling Technology: Publication no. 23, 2014. [24] S. Pelizza, D. Peila, Soil and rock reinforcements in tunnelling, Tunn. Undergr. Space Technol. 8 (1993) 357–372, https://doi.org/10.1016/0886-7798(93)90020- V. [25] P. Lunardi, Design and Construction of Tunnels: Analysis of Controlled Deformation in Rocks and Soils (ADECO-RS), Springer-Verlag, Berlin, Heidelberg, 2008. [26] S. Raschka, V. Mirjalili, Python Machine Learning - Third Edition: Machine learning and deep learning with python, scikit … -learn, and tensorflow 2, Packt Publishing Limited, [S.l.], 2019. [27] R.S. Sutton, A.G. Barto, Reinforcement Learning: An Introduction, Second edition, The MIT Press, Cambridge Massachusetts, 2018. [28] ITA Working Group Conventional Tunnelling, General Report on Conventional Tunnelling Method, Distrib, ITA Secretariat c/o EPFL, Lausanne, 2009. [29] P.A. Vermeer, N. Ruse, T. Marcher, Tunnel heading stability in drained ground, Felsbau 20 (2002) 8–18. [30] M.A. Kanji, Critical issues in soft rocks, J. Rock Mech. Geotech. Eng. 6 (2014) 186–195, https://doi.org/10.1016/j.jrmge.2014.04.002. [31] T. Marcher, S. Stauder, M. Winkler, HSSR – Ein Versuch der Einordnung und Abgrenzung des Materials, in: T. Marcher (Ed.), Beitr¨age zum 1. Hard Soil – Soft Rock (HSSR) Minisymposium: Charakterisierung, Modellierung und experimentelle Untersuchungen von Uebergangsgesteinen – Aktuelles aus Forschung und Entwicklung, Graz, 2020. [32] G. Anagnostou, K. Kov´ari, Face stability conditions with earth-pressure-balanced shields, Tunn. Undergr. Space Technol. 11 (1996) 165–173, https://doi.org/ 10.1016/0886-7798(96)00017-X. [33] A.A. Borovkov, Random walks and factorisation identities, in: A.A. Borovkov (Ed.), Probability Theory, Springer London, London, 2013. [34] C.J.C.H. Watkins, Learning from Delayed Rewards. Ph.D. Thesis, Oxford, 1989. [35] C.J.C.H. Watkins, P. Dayan, Q-learning, Mach. Learn. 8 (1992) 279–292, https:// doi.org/10.1007/BF00992698. [36] A.A. Javadi, M. Rezania, Applications of artificial intelligence and data mining techniques in soil modeling, Geomech. Eng. 1 (2009) 53–74, https://doi.org/ 10.12989/gae.2009.1.1.053. [37] M.A. Shahin, M.B. Jaksa, H.R. Maier, Recent advances and future challenges for artificial neural Systems in Geotechnical Engineering Applications, Adv. Artific. Neural Syst. 2009 (2009) 1–9, https://doi.org/10.1155/2009/308239. [38] H. Kinsley, Reinforcement Learning W/ Python. https://pythonprogramming.net /q-learning-reinforcement-learning-python-tutorial/, 2019. [39] R.H.R. Hahnloser, R. Sarpeshkar, M.A. Mahowald, R.J. Douglas, H.S. Seung, Digital selection and analogue amplification coexist in a cortex-inspired silicon circuit, Nature 405 (2000) 947–951, https://doi.org/10.1038/35016072. [40] X. Glorot, A. Bordes, Y. Bengio, Deep sparse rectifier neural networks, in: G. Gordon, D. Dunson, M. Dudik (Eds.), Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics, 2011, pp. 315–323. [41] Y. Lecun, L. Bottou, Y. Bengio, P. Haffner, Gradient-based learning applied to document recognition, Proc. IEEE 86 (1998) 2278–2324, https://doi.org/10.1109/ 5.726791. [42] D.H. HUBEL, T.N. WIESEL, Shape and arrangement of columns in cat’s striate cortex, J. Physiol. 165 (1963) 559–568, https://doi.org/10.1113/jphysiol.1963. sp007079. [43] F."}], "source_title": "Reinforcement learning based process optimization and strategy development in conventional tunneling", "source_doi": "10.1016/j.autcon.2021.103701", "source_url": "https://www.sciencedirect.com/science/article/pii/S0926580521001527", "source_license": "cc-by", "zotero_item_key": "RXHUVV3X"}
{"id": "12bb525b9ee69532", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Wheel Loader Scooping Controller Using Deep Reinforcement Learning; DOI: 10.1109/access.2021.3056625; license: cc-by-sa."}, {"role": "assistant", "content": "The first involves writing Pulse Width Modulation (PWM) and direction commands to the actuators, whereas the other reads and publishes the current voltage of the linear actuator’s feedback potentiometer. 24148 VOLUME 9, 2021\n[page 5] O. Azulay, A. Shapiro: Wheel Loader Scooping Controller Using Deep Reinforcement Learning V. SCOOPING CONTROLLER ARCHITECTURE In this section, we present our methodology for the DRL scooping motion controller. We describe the implementation of our complete learning system in a virtual environment and its training progression. A. PROBLEM FORMULATION We consider the combined bucket-scooping process of approaching and penetrating the pile, lifting, and then tilt- ing the bucket until breakout using UGV with a loading mechanism. Our goal is to find a closed-loop policy using RL methods that maps sensor observations to robot action. We thus need a policy function at = f (st), where st is a parameterized vector of the current observed robot state using onboard sensor readings and at is the action composed of the lift, tilt, and throttle commands. Given the dynamics of the robot, we limit the linear velocity to 0.1m/s for moving towards and away from the pile, whereas lift and tilt actuators were not limited and could extract at any of their maximum velocities. B. SIMULATION ENVIRONMENT The training procedure of our model was implemented in a virtual 3D environment simulated by a Gazebo. We used a particle-based discrete representation of the pile. This simpli- fication of the real-world surfaces enabled the use of common physics simulators, which can maintain a close resemblance to reality. The number of particles impacts the real-time factor within the simulation considerably, which in turn represents the total run-time. In our environment, each pile of parti- cles contained 72 particles with radii of 3.4 cm and a mass of 0.3kg. Only spherical particles were used in this study. The simulation properties were carefully chosen to mimic the physics of the real robot as well as possible. The robot scooped gravel with a bulk density of 1.3 ton/m3 (bucket capacity is approximately 5[L], 7[kg]). Although the simu- lation was based on a simplified scenario which we found to be sufficient to achieve our goal, other users could vary the environmental properties, such as the pile geometry, number of particles, and the maximum actuator load. In addition, users could choose to define the state according to other features (e.g., object orientation or angular velocities) as well as the authorized actions (e.g., discrete or continuous) for learning purposes. Also, users could build new models to independently collect data using the open-source code pro- vided here [17]. C. STATE REPRESENTATION The observation is the raw information provided by the robot sensors and the state is a compact depiction of this observa- tion that includes the information necessary for the robot to choose its actions. Following the robotic priors in [26], [27], we defined the state using fused sensor measurements as a 12-dimensional vector consisting of the arm link positions (xt), the normalized mass of the excavated earth (wt), the FIGURE 5. Komodo robot with an arm mechanism and a pile of particles, simulated in the Gazebo physics engine. relative distance to the pile (dt), the state of the robot’s DoF (v, θT , θL) and their relative changes (1t). A visual depiction of these parameters is provided in Fig. 5. Deriving the amount of the excavated earth within the simulation environment was done directly using the simulation’s built-in functions. Measuring it in a real-world experiment could be done indi- rectly by approximation using the torque applied to the bucket joint as: τ = VSγ grcm (4) where VS is the excavated earth volume, γ is the material density, g is the gravitational acceleration and rcm is the approximate relative distance between the bucket CoM to the joint. The total force acting on the bucket can be decomposed into three main forces [6]: the shear force, the gravity force, and the remolding force. By deriving the amount of exca- vated earth from the applied torque, we included torque-to- mass extraction of the shearing force as the bucket penetrates the earth. Therefore, we limit our proposed controller to scooping earth with low adhesion, which is applicable to numerous earth-moving scenarios, since the main purpose of the wheel loader is the transportation of already excavated material. FIGURE 6."}], "source_title": "Wheel Loader Scooping Controller Using Deep Reinforcement Learning", "source_doi": "10.1109/access.2021.3056625", "source_url": "https://ieeexplore.ieee.org/ielx7/6287639/9312710/09344588.pdf", "source_license": "cc-by-sa", "zotero_item_key": "IDZBIHT7"}
{"id": "01a08dd9917c28e3", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Reinforcement learning based process optimization and strategy development in conventional tunneling; DOI: 10.1016/j.autcon.2021.103701; license: cc-by."}, {"role": "assistant", "content": "Chollet (Ed.), Deep Learning with Python, Manning, Shelter Island, NY, 2018. [44] M. Abadi, A. Agarwal, P. Barham, E. Brevdo, Z. Chen, C. Citro, G.S. Corrado, A. Davis, J. Dean, M. Devin, S. Ghemawat, I. Goodfellow, A. Harp, G. Irving, M. Isard, Y. Jia, R. Jozefowicz, L. Kaiser, M. Kudlur, J. Levenberg, D. Man´e, R. Monga, S. Moore, D. Murray, C. Olah, M. Schuster, J. Shlens, B. Steiner, I. Sutskever, K. Talwar, P. Tucker, V. Vanhoucke, V. Vasudevan, F. Vi´egas, O. Vinyals, P. Warden, M. Wattenberg, M. Wicke, Y. Yu, X. Zheng, TensorFlow: Large-Scale Machine Learning on Heterogeneous Systems. https://www.tensor flow.org/, 2015. [45] F. Chollet, Others, Keras, 2015. [46] ¨Osterreichisches Normungsinstitut, Untertagebauarbeiten: Teil 1: Zyklischer Vortrieb, Wien 91.010.20; 93.020, 2019. [47] Z. Zhao, Y. Liang, X. Jin, Handling large-scale action space in deep Q network, in: 2018 International Conference on Artificial Intelligence and Big Data (ICAIBD), IEEE, Chengdu, 2018, pp. 93–96. [48] T. Zahavy, M. Haroush, N. Merlis, D.J. Mankowitz, S. Mannor, Learn what not to learn: action elimination with deep reinforcement learning, in: Advances in Neural Information Processing Systems 31 (NIPS 2018), Montreal, Canada, 2018, pp. 3566–3577. [49] ¨OGG, Guideline for the Geotechnical Design of Underground Structures with Conventional Excavation: Ground characterization and coherent procedure for the determination of excavation and support during design and construction. Translated from version 2.1, second.first, Salzburg, 2010. [50] H. Lauffer, Gebirgsklassifizierung für den Stollenbau, Geol. Bauwesen 24 (1958) 46–51. [51] P. Zhang, H. Li, Q.P. Ha, Z.-Y. Yin, R.-P. Chen, Reinforcement learning based optimizer for improvement of predicting tunneling-induced ground responses, Adv. Eng. Inform. 45 (2020) 101097, https://doi.org/10.1016/j.aei.2020.101097. [52] J. van Eldert, H. Schunnesson, D. Johansson, D. Saiang, Application of measurement while drilling technology to predict rock mass quality and rock support for Tunnelling, Rock Mech. Rock. Eng. 53 (2020) 1349–1358, https://doi. org/10.1007/s00603-019-01979-2. [53] A. Radinger, F. Fasching, G. Pack, I. Kreutzer, D. Kostial, Consistent exploration by probe drilling and TSWD through the example of the Koralm tunnel/Konsequente Vorauserkundung mittels Bohrungen und TSWD am Beispiel des Koralmtunnels, Geomechanik Tunnelbau 7 (2014) 540–550, https://doi.org/10.1002/ geot.201400038. G.H. Erharter et al."}], "source_title": "Reinforcement learning based process optimization and strategy development in conventional tunneling", "source_doi": "10.1016/j.autcon.2021.103701", "source_url": "https://www.sciencedirect.com/science/article/pii/S0926580521001527", "source_license": "cc-by", "zotero_item_key": "N5Q33JID"}
{"id": "d4efa3c3884b0d2f", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Wheel Loader Scooping Controller Using Deep Reinforcement Learning; DOI: 10.1109/access.2021.3056625; license: cc-by-sa."}, {"role": "assistant", "content": "FIGURE 12. Velocity, lift and tilt motion recorded in 5 consecutive scooping cycles. which thus illustrates the sensitivity of the resistive model to contacts. We also conducted another experiment to test the con- troller’s adaptability to different environments, but with typi- cal features such as different inclination angles of the terrain. A detailed comparison of the bucket tip trajectories (3 trials for each inclination angle) over a scooping cycle can be seen in Fig. 13 with three different pile inclination angles. Steeper pile inclination angles of the earth caused higher resistance force from the pile, leading to an earlier exit phase initialization, as can be seen in Fig 13. Overcoming different inclination angles supports the claim that our DRL-based controller can perform successfully in arbitrarily designed piles. The results of both experiments suggest that the tran- sition between different phases of the scooping process does FIGURE 13. Bucket tip trajectories over 3 different mean earth inclination angles (A). Applied torque at the bucket joint (B). FIGURE 14. Comparisons between the scooping process in the simulations and real-life experiments. TABLE 2. Comparison of scooping experiments. not depend on a single feature of the state, but rather on a combination of input features. The controller continued the filling phase until reaching the torque and depth thresholds, leading to an exit phase initialization, which appeared to be similar across all the attempts, as can be seen in Fig. 13. Comparing the DRL con- troller performance to the manually driven scooping attempts revealed differences between various processes of the scoop- ing cycle, as can be seen in Fig. 14. Due to the fact that we are not expert operators, it was extremely challenging to main- tain a successful smooth, fast continuous actions. The DRL controller achieved smoother action transitions and faster scooping cycles (average manual operation time was 10[sec], compared to 6 [sec] for the DRL-based controller). However, our manual attempts resulted in slightly better bucket weights, where the manual control resulted in an 24152 VOLUME 9, 2021\n[page 9] O. Azulay, A. Shapiro: Wheel Loader Scooping Controller Using Deep Reinforcement Learning average of 6.1[kg] and the DRL controller resulted in an average of 5.5[kg] (9% difference), as shown in Table 2. Comparing the experimental results of our DRL controller to the simulation polices also pointed to differences between the phases of the scooping cycle. These included a difference between the penetration angle, that reached a maximum of 7.5 ± 1.0 [deg], higher loading cycles in the experiment, and a different trajectory within the second phase of the cycle, as illustrated in Fig. 14. These can be explained by the dis- crepancies between the simulation environment and the real- life experiments, such as particles vs. earth, action execution time, and mass approximation. VII. CONCLUSION In this article, a machine learning-based deep reinforce- ment learning scooping motion controller was applied for UGV with a custom-built scooping mechanism to perform scooping cycles with three DoF (lift, tilt, velocity) by fus- ing 12-dimensional sensor data as inputs to output actuator commands. The controller was trained using the actor-critic, off-policy Deep Deterministic Policy Gradient (DDPG) algo- rithm without any user-provided demonstrations. The learn- ing of the policy network is accomplished in a simplified simulation scenario, where the soil was represented as parti- cles. The scooping policy converged at the end of the training phase with an average loading cycle time of 5 seconds and an average fill factor of 65% from full bucket capacity with an overfill. To investigate the performance and adaptability of our scooping controller, we tested our controller in several experiments with three different mean inclination angles of the earth. Overall, the DRL-based control exhibited good performance in terms of both achieved visual bucket fill with varying scooped earth weights of 4.1 −7.2[kg], and 5.1 −7.1[sec] cycle time. Comparisons between the man- ual and our DRL controller indicated that human control led to the highest average bucket load. Although the DRL controller may not yet be on a par with human control for the highest volume, the DRL controller here did achieve smoother trajectories and lower cycle times than the man- ual control."}], "source_title": "Wheel Loader Scooping Controller Using Deep Reinforcement Learning", "source_doi": "10.1109/access.2021.3056625", "source_url": "https://ieeexplore.ieee.org/ielx7/6287639/9312710/09344588.pdf", "source_license": "cc-by-sa", "zotero_item_key": "IDZBIHT7"}
{"id": "1af750dbb96012ca", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Reinforcement learning based process optimization and strategy development in conventional tunneling; DOI: 10.1016/j.autcon.2021.103701; license: cc-by."}, {"role": "assistant", "content": "We give detailed lists of the possible actions in Section 3.2.1 and explain the exact content of the state and the rewards in the Sections 3.3.1 and 3.3.2 respectively. Fig. 1. Schematic diagram of cyclical and sequential processes in conventional tunnel excavation. The size of loops represents a qualitative hierarchy with smaller loops being components of bigger loops. The excavation sequence decision – loop is given in bold, as this will be the main focus of this paper. G.H. Erharter et al.\n[page 3] Automation in Construction 127 (2021) 103701 3 2.1. Simplifications As shown in Fig. 1, conventional tunneling is a complex process that consists of several subprocesses which again have subprocesses etc. Attempting to formulate this whole system as a RL problem with all involved details is out of the scope of an initial study which should serve as the basis for future research. Additionally, it must also be considered that RL is still at the very beginning of practical applications. Conse­ quently, we had to take several simplifications of conventional tunneling to reduce the general complexity of the simulation and the size of the state - and action space. Still, we consider modelling the major decision- loop and the actions and rewards, not to be too far from a realistic tunnel approach, and will contribute with important insights to the use of RL in the optimization of the conventional tunnel cycle. Some important simplifications in comparison to reality were made in this study: ▪ Available partial excavation methods are restricted to either top heading - or a combined bench and invert excavation (described as “bench excavation” hereafter) with a specific tunnel geometry. ▪ Already excavated parts are considered to be stable and the process of tunnel lining installation is not dealt with in the simulation. ▪ If installed, support ahead of the face always consists of 10 m long face anchors. The supported area ahead is stable and excavation within that stable area cannot lead to failure. No other types of rock support are considered, such as radial bolts and shotcrete. ▪ We evaluate if stable conditions are at hand in the excavation area by the face pressure equation for open face tunneling after [29] (see Eq. (3) in Section 3.3.2). We chose this analytical solution as it is a computationally efficient way for a stability assessment that tells if stability is given or not at a certain ground type. Although this is only one aspect of the tunnel stability considerations, we see this approach as sufficient for the present initial study. ▪ The stability assessment considers the cross-sectional area only and no longitudinal effects. ▪Rockmass quality is reduced to few mechanical values repre­ senting “favorable” or “unfavorable” rockmass conditions. ▪ There are only two available advance lengths (i.e. the length of one blasting round/round of excavation) with 2 and 4 m each. ▪ Information from probe drilling or deformation monitoring is not simulated and used in the decision process. That means that the agent has no information about the rockmass ahead of the face or no information about eventual deformation behind the face. 3. A simulation of conventional tunneling As the main goal of this study is to train a RL agent to execute an excavation sequence as efficiently as possible, we designed a simulation of such a scenario. The simulation consists of a longitudinal tunnel section of a specific length (tl), where two different types of ground conditions can occur – one favorable and one unfavorable. Before the excavation, the agent is unaware of the distribution of ground types and the distribution of the ground types is only revealed by the excavation itself. The ultimate goal of the agent is to achieve a breakthrough of both the top heading and the bench of the tunnel or in other words, the po­ sition of the top heading excavation (posth) and the position of the bench excavation (posbi) must be greater than, or equal to tl. During the exca­ vation the agent can choose from different actions, e.g. top heading excavation with 2 m advance length, bench excavation with 4 m advance length and installation of face support etc. 3.1. Geotechnical scenario For this study’s simulation we have chosen the following tunneling scenario: The total length of the tunnel (tl) is 200 m, as this does not lead to an excessively large state space in the RL model (see Section 3.3.1) but is still a realistic length."}], "source_title": "Reinforcement learning based process optimization and strategy development in conventional tunneling", "source_doi": "10.1016/j.autcon.2021.103701", "source_url": "https://www.sciencedirect.com/science/article/pii/S0926580521001527", "source_license": "cc-by", "zotero_item_key": "RXHUVV3X"}
{"id": "9f844086265a3f3d", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Cutterhead and Cutting Tools Configurations in Coarse Grain Soils; DOI: 10.2174/1874836801711010182; license: cc-by."}, {"role": "assistant", "content": "The single disc cutters have a single point of contact at the face and tend to skid or be stagnant in the soft or not compact soils, wearing flat on one side and making excavation of the over-sized particles difficult. Frequent work of changing cutting tools at more time and expenses has to be executed during the whole tunnel construction. Whereas, disc cutters with tungsten carbide inserts can excavate cobbles and boulders with a higher efficiency because they have multiple points of contact at the face. The multiple buttons act like traction, allowing the carbide cutters to ‘grip’ in soft or loosened ground and keep rotating. As compared to performances of the twin disc cutters, it was found that the twin disc cutters were more easily worn out than single disc cutters when building a railway tunnel in Beijing sandy gravel/cobble soils with boulders [32]. So, it is recommended that the plate type cutterheads equipped with the chip inserted single disc cutters in combination with the more aggressive soft ground tools are adopted when boring tunnels in the water-rich coarse grain soils. Fig. (20). Clogging at the cutterhead center. 4.4. Reasonable Movement of Shield Machines As cutting tools move in the excavation face they are taking up highly viable loads. The wear and failure of the tools is the result of such cyclic loadings. As revealed in Figs. (15, 17), dropping off of tips and fractures of cutting tools are prevailing damages caused by cobbles or boulders when they abruptly hit these over-sized particles. To mitigate the adverse effects, deciding the slow and reasonable movements (advance, rotation) of a shield machine is of great importance. Particularly for hard boulder-laden ground when using disc cutters the penetration rate per cutterhead rotation should be well regulated and restricted to avoid that boulders will be ripped-out of the soil matrix, endangering the stability of the cutting face and damaging cutterhead and cutting tools. For a spoke type cutterhead shield machine, the well planned sensible and slow movements for the specific ground conditions can greatly reduce the damages to precut bits as well as to scrapers caused by collisions with big cobbles or boulders. 4.5. Robust Shield Machines Again Wears The wears and damages to cutting tools, cutterhead and other components do exist if shield tunneling will be performed in the coarse grain soils. Particularly, when a long tunnel will be constructed, it is absolutely necessary to take whatever, state-of-the-art measures to reduce the adverse influences. In recent years, large improvements have been achieved in relation to the cutting tool durability and wear protection on cutterhead as well as other components of a shield machine [33]. The highlights are more specialized, robust wear resistant tools. These robust tools, such as pressure compensated disc cutters of monobloc design including options with carbide inserts, soft ground tools (scrapers, rippers) using carbide cutting edges or inserts in combinations with hardfacing, will allow for a longer lifespan of the cutting tools under the coarse grain ground conditions. Using lately developed twin layer plating materials, overall replaceable wear protections of the cutterhead in weld-on or bolted-on designs can also be effectively applied to the shield machines in coarse grain soils [34, 35]. The designs consist of wedges to protect the cutters, wear\n[page 17] Cutterhead and Cutting Tools Configurations The Open Construction and Building Technology Journal, 2017, Volume 11 197 plates for the face, wear protection at the cutterhead periphery, grill bars to protect the cutterhead periphery, and grill bars to protect the scrapers from big blocks and control the maximum block size to enter the cutterhead. 4.6. Interventions for Changing Tools Interventions and chamber access for repairing tools and other components are sometimes indispensable to some demanding and challenging works of the tunnel construction in coarse grain soils. Compared with changing cutting tools under compressed air condition, changing tools in free air, or “atmospheric cutter change” based on a hollow cutterhead structure for a large diameter shield machine [33, 35], is a better choice."}], "source_title": "Cutterhead and Cutting Tools Configurations in Coarse Grain Soils", "source_doi": "10.2174/1874836801711010182", "source_url": "https://openconstructionbuildingtechnologyjournal.com/VOLUME/11/PAGE/182/PDF/", "source_license": "cc-by", "zotero_item_key": "7KDIA2X2"}
{"id": "327fc5ea5d0bcae9", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Detection of Ionospheric Scintillation Based on XGBoost Model Improved by SMOTE-ENN Technique; DOI: 10.3390/rs13132577; license: cc-by-sa."}, {"role": "assistant", "content": "[page 1] remote sensing Article Detection of Ionospheric Scintillation Based on XGBoost Model Improved by SMOTE-ENN Technique Mengying Lin 1 , Xuefen Zhu 1,*, Teng Hua 1, Xinhua Tang 1, Gangyi Tu 2 and Xiyuan Chen 1 \u0001\u0002\u0003\u0001\u0004\u0005\u0006\u0007\b \u0001\u0002\u0003\u0004\u0005\u0006\u0007 Citation: Lin, M.; Zhu, X.; Hua, T.; Tang, X.; Tu, G.; Chen, X. Detection of Ionospheric Scintillation Based on XGBoost Model Improved by SMOTE-ENN Technique. Remote Sens. 2021, 13, 2577. https://doi.org/ 10.3390/rs13132577 Academic Editor: Michael E. Gorbunov Received: 1 June 2021 Accepted: 29 June 2021 Published: 1 July 2021 Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affil- iations. Copyright: © 2021 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/). 1 Key Laboratory of Micro-Inertial Instrument and Advanced Navigation Technology of Ministry of Education, Sipailou Campus, School of Instrument Science and Engineering, Southeast University, Nanjing 210096, China; linmengying@seu.edu.cn (M.L.); 220203499@seu.edn.cn (T.H.); xinhua.tang@seu.edu.cn (X.T.); chxiyuan@seu.edu.cn (X.C.) 2 School of Electronic and Information Engineering, Nanjing University of Information Science and Technology, Ning Liu Road, Nanjing 210044, China; tugangyi@nuist.edu.cn * Correspondence: zhuxuefen@seu.edu.cn; Tel.: +86-136-4516-1372 Abstract: Ionospheric scintillation frequently occurs in equatorial, auroral and polar regions, posing a threat to the performance of the global navigation satellite system (GNSS). Thus, the detection of ionospheric scintillation is of great significance in regard to improving GNSS performance, es- pecially when severe ionospheric scintillation occurs. Normal algorithms exhibit insensitivity in strong scintillation detection in that the natural phenomenon of strong scintillation appears only occasionally, and such samples account for a small proportion of the data in datasets relative to those for weak/moderate scintillation events. Aiming at improving the detection accuracy, we proposed a strategy combining an improved eXtreme Gradient Boosting (XGBoost) algorithm by using the synthetic minority, oversampling technique and edited nearest neighbor (SMOTE-ENN) resampling technique for detecting events imbalanced with respect to weak, medium and strong ionospheric scintillation. It outperformed the decision tree and random forest by 12% when using imbalanced training and validation data, for tree depths ranging from 1 to 30. For different degrees of imbalance in the training datasets, the testing accuracy of the improved XGBoost was about 4% to 5% higher than that of the decision tree and random forest. Meanwhile, the testing results for the improved method showed significant increases in evaluation indicators, while the recall value for strong scintillation events was relatively stable, above 90%, and the corresponding F1 scores were over 92%. When testing on datasets with different degrees of imbalance, there was a distinct increase of about 10% to 20% in the recall value and 6% to 11% in the F1 score for strong scintillation events, with the testing accuracy ranging from 90.42% to 96.04%. Keywords: GNSS; ionospheric scintillation detection; XGBoost; SMOTE-ENN 1. Introduction The ionosphere, the atmosphere at about 60 to 1000 km from the ground, is modulated by the ionizing effects of solar radiation, particle precipitation and the geomagnetic field. There are some typical ionospheric phenomena, such as the equatorial ionization anomaly (EIA) and equatorial plasma bubbles at low attitudes [1–3], as well as the tongue of ion- ization at high latitudes [4,5], by which radio waves such as global navigation satellite system (GNSS) signals may be severely affected. When passing through the ionospheric irregularities, the signals are plagued with rapid fluctuation, phase shifts, delay, multi- path, and even loss of tracking loop. More seriously, this phenomenon of ionospheric scintillation occurs more frequently and severely in both low-latitude regions and polar regions, compromising positioning accuracy and continuity [6,7]. In high-latitude regions, the occurrence of ionospheric scintillation is more apparent during geomagnetic storms, and the formation of irregular structures and ionospheric scintillation activities appears to Remote Sens."}], "source_title": "Detection of Ionospheric Scintillation Based on XGBoost Model Improved by SMOTE-ENN Technique", "source_doi": "10.3390/rs13132577", "source_url": "https://www.mdpi.com/2072-4292/13/13/2577/pdf?version=1625212202", "source_license": "cc-by-sa", "zotero_item_key": "4YEIWMK3"}
{"id": "e1d3c66a098963c0", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Reinforcement learning based process optimization and strategy development in conventional tunneling; DOI: 10.1016/j.autcon.2021.103701; license: cc-by."}, {"role": "assistant", "content": "We give detailed lists of the possible actions in Section 3.2.1 and explain the exact content of the state and the rewards in the Sections 3.3.1 and 3.3.2 respectively. Fig. 1. Schematic diagram of cyclical and sequential processes in conventional tunnel excavation. The size of loops represents a qualitative hierarchy with smaller loops being components of bigger loops. The excavation sequence decision – loop is given in bold, as this will be the main focus of this paper. G.H. Erharter et al.\n[page 3] Automation in Construction 127 (2021) 103701 3 2.1. Simplifications As shown in Fig. 1, conventional tunneling is a complex process that consists of several subprocesses which again have subprocesses etc. Attempting to formulate this whole system as a RL problem with all involved details is out of the scope of an initial study which should serve as the basis for future research. Additionally, it must also be considered that RL is still at the very beginning of practical applications. Conse­ quently, we had to take several simplifications of conventional tunneling to reduce the general complexity of the simulation and the size of the state - and action space. Still, we consider modelling the major decision- loop and the actions and rewards, not to be too far from a realistic tunnel approach, and will contribute with important insights to the use of RL in the optimization of the conventional tunnel cycle. Some important simplifications in comparison to reality were made in this study: ▪ Available partial excavation methods are restricted to either top heading - or a combined bench and invert excavation (described as “bench excavation” hereafter) with a specific tunnel geometry. ▪ Already excavated parts are considered to be stable and the process of tunnel lining installation is not dealt with in the simulation. ▪ If installed, support ahead of the face always consists of 10 m long face anchors. The supported area ahead is stable and excavation within that stable area cannot lead to failure. No other types of rock support are considered, such as radial bolts and shotcrete. ▪ We evaluate if stable conditions are at hand in the excavation area by the face pressure equation for open face tunneling after [29] (see Eq. (3) in Section 3.3.2). We chose this analytical solution as it is a computationally efficient way for a stability assessment that tells if stability is given or not at a certain ground type. Although this is only one aspect of the tunnel stability considerations, we see this approach as sufficient for the present initial study. ▪ The stability assessment considers the cross-sectional area only and no longitudinal effects. ▪Rockmass quality is reduced to few mechanical values repre­ senting “favorable” or “unfavorable” rockmass conditions. ▪ There are only two available advance lengths (i.e. the length of one blasting round/round of excavation) with 2 and 4 m each. ▪ Information from probe drilling or deformation monitoring is not simulated and used in the decision process. That means that the agent has no information about the rockmass ahead of the face or no information about eventual deformation behind the face. 3. A simulation of conventional tunneling As the main goal of this study is to train a RL agent to execute an excavation sequence as efficiently as possible, we designed a simulation of such a scenario. The simulation consists of a longitudinal tunnel section of a specific length (tl), where two different types of ground conditions can occur – one favorable and one unfavorable. Before the excavation, the agent is unaware of the distribution of ground types and the distribution of the ground types is only revealed by the excavation itself. The ultimate goal of the agent is to achieve a breakthrough of both the top heading and the bench of the tunnel or in other words, the po­ sition of the top heading excavation (posth) and the position of the bench excavation (posbi) must be greater than, or equal to tl. During the exca­ vation the agent can choose from different actions, e.g. top heading excavation with 2 m advance length, bench excavation with 4 m advance length and installation of face support etc. 3.1. Geotechnical scenario For this study’s simulation we have chosen the following tunneling scenario: The total length of the tunnel (tl) is 200 m, as this does not lead to an excessively large state space in the RL model (see Section 3.3.1) but is still a realistic length."}], "source_title": "Reinforcement learning based process optimization and strategy development in conventional tunneling", "source_doi": "10.1016/j.autcon.2021.103701", "source_url": "https://www.sciencedirect.com/science/article/pii/S0926580521001527", "source_license": "cc-by", "zotero_item_key": "N5Q33JID"}
{"id": "3c233aa914412911", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Reinforcement learning based process optimization and strategy development in conventional tunneling; DOI: 10.1016/j.autcon.2021.103701; license: cc-by."}, {"role": "assistant", "content": "[page 1] Automation in Construction 127 (2021) 103701 Available online 17 April 2021 0926-5805/© 2021 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/). Reinforcement learning based process optimization and strategy development in conventional tunneling Georg H. Erharter a,*, Tom F. Hansen b, Zhongqiang Liu b, Thomas Marcher a a Graz University of Technology, Institute for Rock Mechanics and Tunnelling, Rechbauerstraße 12, Graz, Austria b Norwegian Geotechnical Institute, Oslo, Norway A R T I C L E I N F O Keywords: Conventional tunneling Reinforcement learning Tunnel excavation strategy Machine learning Excavation sequences A B S T R A C T Reinforcement learning (RL) - a branch of machine learning - refers to the process of an agent learning to achieve a certain goal by interaction with its environment. The process of conventional tunneling shows many similar­ ities, where a geotechnician (agent) tries to achieve a breakthrough (goal) by excavating the rockmass (envi­ ronment) in an optimum way. In this paper we present a novel RL based framework for strategy development for conventional tunneling. We developed a virtual environment with the goal of a tunnel breakthrough and with a deep Q-network as the agent’s architecture. It can choose from different excavation sequences to reach that goal and learns to do so in an economical and safe way by getting feedback from a specially designed reward system. Result analyses show that the optimal policies have great similarities to current practices of sequential tunneling and the framework has the potential to discover new tunneling strategies. 1. Introduction Digitalization in tunneling is an ongoing process including topics like Artificial Intelligence (AI) technology [1], Building Information Modelling [2] or embedded strain measurements in shotcrete lining [3]. Machine Learning (ML) applications in particular have so far mostly focused on classification tasks based on supervised ML: e.g. [4–6] apply artificial neural networks (ANN) for rockmass behavior classification of tunnel boring machine (TBM) data; [7] use supervised learning methods for automatic work progress identification in NATM (New Austrian Tunneling Method) tunneling; see [8–11] for state-of-the-art reviews on this topic. Applications of unsupervised ML are fewer in number and often related to the search for less dependency on labelled datasets or more objectiveness (e.g. [12,13]). By the time of this writing reinforcement learning (RL) is mostly a matter of research, but already shows mature problem-solvers for game like scenarios [14,15]. Currently, there is a transition from academia to real-world prototypes, with RL-examples like the optimization of a manufacturing process [16], for real-time steering of hydrocarbon drilling [17,18], in optimization of power grids [19] and control systems in general, for AUVs [20], and in robotics and other autonomous vehi­ cles. To our knowledge, however, there is no published application of RL to geotechnics in general or tunneling in particular. Today’s conventional tunneling – sometimes referred to as “drill and blast tunneling” or “sequential excavation method” – is the classical way of tunnel construction and is the product of more than a century of engineering experience [21]. Great flexibility to adapt to changing ground conditions is one of the main benefits of this type of tunneling, but technical and economic success of the excavation is dependent on the experience of the involved engineers and workers. Albeit experience is undoubtedly valuable, depending on it sometimes goes along with simple repetition of “proven ways” or even negligence of innovation. Furthermore, developments in conventional tunneling often have a strong connection to their nation of origin (e.g. Austria: New Austrian Tunnelling Method [22], Norway: Norwegian Tunneling Method [23], Italy: New Italian Tunneling Method [24,25], etc.) which raises concerns about biased researchers and engineers. The goal of this study is to take a first step in the direction of a conceptual RL-model for optimum decision making in conventional tunneling that is as free as possible of conser­ vatism and national biases. Furthermore, as shown in the development of the RL-agent AlphaGO [15], RL systems have the potential to find new solutions to old problems and thus discover unimagined strategies."}], "source_title": "Reinforcement learning based process optimization and strategy development in conventional tunneling", "source_doi": "10.1016/j.autcon.2021.103701", "source_url": "https://www.sciencedirect.com/science/article/pii/S0926580521001527", "source_license": "cc-by", "zotero_item_key": "N5Q33JID"}
{"id": "b3e4e5ce02060f5c", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Two-phase multi-model automatic brain tumour diagnosis system from magnetic resonance images using convolutional neural networks; DOI: 10.1186/s13640-018-0332-4; license: cc-by-sa."}, {"role": "assistant", "content": "2 The architecture of the CNN used in this study. The name of each network layer is provided. The figure shows a snapshot of the whole network architecture To facilitate the convolution process, the dimensions of the convolutional kernel and the CNN input should be matched. If the CNN is a multidimensional array, then the kernel should be a multidimensional array as well. If the CNN input is a two-dimensional image [38], a two-dimensional kernel w is used, as presented in Eq. 1. The output of the convolutional layers, i.e., layers 2 and 6, will apply a sigmoidal activation function to the obtained convolutional output y in Eq. 1 to provide the new output f modelled as function of the input y with Eqs. 2 or 3: f (y) = tanh(y) (2) f (y) = 1 1 + e−y (3) The output of the ReLU layers, i.e., layers 3 and 7, will apply a linear activation function to the neuron output y as shown in Eq. 4: f (y) = max(0, y) (4) Layers 4 and 8 are the cross-channel normalization lay- ers with five channels. The 3 × 3 max-pooling layers are layers 5 and 9. These layers are very simple because they take a K ×K region and output a single value, and then the output is an N K × N K layer. Layer 9 provides the extracted features in a dimension of 1 × 43264, which are then fed to the SVM to carry out the image classification. 3.1.3 Error-correcting output codes support vector machines (ECOC-SVMs) Dietterich and Bakiri presented an error-correcting out- put codes (ECOC) approach to correct the data error when input enters into a channel. The approach uses error-correcting codes to convert the binary classifiers to a multi-class classification [39]. It has been successfully used in many applications, such as face recognition [40], face verification [41], text recognition [42], and digital manuscript classification [43]. Error-correcting output codes support vector machines (ECOC-SVMs) is a method that combines the SVMs and ECOC. It was proposed to solve the online identification and feature extraction problems. When applying ECOC- SVMs to multi-classification, the k classes classification is transformed to the L classes classification by the ECOC matrix. ECOC-SVM functions according to the following steps [44]: 1 Training: The classes are represented in two subsets (0 or 1) for each row of the coding matrix using the SVM binary partition. The SVM takes the row with a value of 0 as the first class and the row with a value of 1 as the second class. The corresponding class assigns a codeword for each column of the matrix. All of the L classes are trained according to rows of the ECOC matrix, where L is the dimensionality of the new feature space and the length of the codeword. 2 Testing: The test data is fed into the learned classifier. The L classifiers recognize the samples and generate output vector Z. The hamming distance\n[page 6] Abd-Ellah et al. EURASIP Journal on Image and Video Processing (2018) 2018:97 Page 6 of 10 Fig. 3 Representative images from the MRIs database that was used in the tumour detection phase. Normal MRIs without brain tumours are in the top row, while abnormal images with brain tumours present are in the bottom row between each vector in the output vector matrix and code matrix is shown in Eq. 5. The classifier chooses the class depending on the minimal distance. ci = argmind(Z, Hi) = L \u0002 j=1 |Zj −Hi,j|, i = 1, 2, ..., k (5) Table 3 The obtained number of features, the consumed feature extraction time, and the accuracy of the different CNN layers Layer number Evaluation metrics No. of features Feature extraction time Accuracy (%) Layer 2 (conv) 290400 0.24 69.16 Layer 3 (ReLU) 290400 0.25 84.58 Layer 4 (norm) 290400 0.33 68.28 Layer 5 (pool) 69984 0.35 66.96 Layer 6 (conv) 186624 0.78 69.16 Layer 7 (ReLU) 186624 0.80 48.01 Layer 8 (norm) 186624 0.83 77.97 Layer 9 (pool) 43264 0.85 99.55 Layer 10 (conv) 64896 1.20 49.33 Layer 11 (ReLU) 64896 1.21 81.93 Layer 12 (conv) 64896 1.40 65.63 Layer 13 (ReLU) 64896 1.41 98.23 Layer 14 (conv) 43264 1.59 88.54 Layer 15 (ReLU) 43264 1.60 96.47 The table indicates that the optimum number of layers with respect to the consumed time and the achieved accuracy is 9 where H is the code matrix with i rows and j columns. 3.2 Brain tumour localization phase Since the BraTS 2013 database includes 3D MRIs, we used its MRIs slice by slice as inputs for the CNN. Our method handles each slice as a 2D image [29], and T2 images were selected only with segmented images."}], "source_title": "Two-phase multi-model automatic brain tumour diagnosis system from magnetic resonance images using convolutional neural networks", "source_doi": "10.1186/s13640-018-0332-4", "source_url": "https://jivp-eurasipjournals.springeropen.com/track/pdf/10.1186/s13640-018-0332-4", "source_license": "cc-by-sa", "zotero_item_key": "2ZZSU54R"}
{"id": "74f7b48fd4f06814", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Reinforcement learning based process optimization and strategy development in conventional tunneling; DOI: 10.1016/j.autcon.2021.103701; license: cc-by."}, {"role": "assistant", "content": "The tunnel’s cross section has a total area of 91.31m2 and a height of 10 m with 58.56m2 and 32.75m2 being the areas of the top heading - and the bench and invert excavation respec­ tively (see Fig. 3). From these areas, equivalent diameters (D, i.e. the diameter of a circle with that area) of 8.41 m and 6.46 m can be computed for both parts of the excavation. There is a penalty if the distance between the top heading’s- and the bench’s tunnel face is too big. This distance (distmax) is set to be 50 m and the idea behind this is, that in many real projects – especially in soil conditions and long tunnels -, the top heading cannot be driven indefi­ nitely long ahead of the bench, because of safety reasons, necessities to Fig. 2. Simplified schematic plot of components of drill and blast tunneling (grey) as a reinforcement learning process - TunnRL. Fig. 3. Tunnel cross section of the given simulation. G.H. Erharter et al.\n[page 4] Automation in Construction 127 (2021) 103701 4 have a fast final ring closure, excavate cross cuts or general construction logistics. We defined two ground types (gt) for the simulation and the relevant ground properties are given in Table 1. Both are considered to be of homogeneous, isotropic and continuous nature and one can imagine them as a type of hard soil/soft rock (HSSR) material [30,31]. With the chosen parameters, gt1 represents “unfavorable” ground conditions and gt2 represents “favorable” ground conditions. The condition for a gt to be favorable/unfavorable is based on the ground properties in combi­ nation with the given tunnel geometry (see above) as evaluated by the chosen stability assessment criterion (see Section 3.3.2). The simulated tunnel is situated above the groundwater table and the permeability is set to 10−5 m/s for both gt. Choosing the given properties to define the gt, is closely connected to the way the stability assessment of the excavation is done (see Eq. (3) in Section 3.3.2). The permeability of 10−5 justifies the use of Eq. (3) for a stability assessment as according to [32] conditions are considered to be drained when the permeability is above 10−7 to 10−6 m/s. Nevertheless, if ground types with lower permeability/undrained conditions are to be used for the simulation, appropriate solutions for stability assessments must be chosen. Implementing more sophisticated stability assessments that take phenomena like ground water conditions into account is desirable but must be done with care as this heavily influences the overall performance of the tunneling simulation and the RL itself. The outlook and discussion of stability assessments in Section 7 goes into more detail on this topic. Using random walks with barriers [33], we created unique, 210 m long geological sections with a decimeter resolution (see Section 3.3.1 for why the sections are 210 and not 200 m long). We created a ground type-vector (gt-vector) of 2100 datapoints by scaling the random walk between 0 and 1, rounding to full numbers and using 0 as gt1 and 1 as gt2 (see Fig. 4 bottom row). To transform the gt-vector to a full geological section with one row for the top heading and one for the bench excavation, the vector is horizontally duplicated to an array of 2 × 2100 datapoints. + 1 is then added to the array so that the number 1 represents gt1, number 2 gt2 and number 0 represents the unexcavated part of the tunnel. The top row of Fig. 4 shows a visualization of such an array, where posth and posbi are at 165 and 125 m respectively. Given the complexity of simulating the process of conventional tunneling (see Section 2), we chose not to complicate the scenario by introducing more gt. However, increasing the number of gt can easily be done with the above described random walk based approach. For example, if it was necessary to simulate four gt then the values of the random walk must be split into four within the boundaries: gt1 < 0.25, 0.25 ≤gt2 < 0.5, 0.5 ≤gt3 < 0.75 and 0.75 ≥gt4. 3.2. Agent Translating the above described geotechnical scenario to RL, the geotechnician who observes the state of the construction and rockmass behavior and makes decisions based on this information, now becomes the “RL agent”. Due to the state-complexity of the problem, we chose a deep Q-network (DQN) as the RL agent."}], "source_title": "Reinforcement learning based process optimization and strategy development in conventional tunneling", "source_doi": "10.1016/j.autcon.2021.103701", "source_url": "https://www.sciencedirect.com/science/article/pii/S0926580521001527", "source_license": "cc-by", "zotero_item_key": "N5Q33JID"}
{"id": "6c3df1621090e661", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Intelligent Classification of Surrounding Rock of Tunnel Based on 10 Machine Learning Algorithms; DOI: 10.3390/app12052656; license: cc-by-sa."}, {"role": "assistant", "content": "[page 1] \u0001\u0002\u0003\u0001\u0004\u0005\u0006\u0007\b \u0001\u0002\u0003\u0004\u0005\u0006\u0007 Citation: Zhao, S.; Wang, M.; Yi, W.; Yang, D.; Tong, J. Intelligent Classification of Surrounding Rock of Tunnel Based on 10 Machine Learning Algorithms. Appl. Sci. 2022, 12, 2656. https://doi.org/10.3390/ app12052656 Academic Editors: Mingfeng Lei, Chenjie Gong and Xianda Shen Received: 7 February 2022 Accepted: 27 February 2022 Published: 4 March 2022 Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affil- iations. Copyright: © 2022 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/). applied sciences Article Intelligent Classification of Surrounding Rock of Tunnel Based on 10 Machine Learning Algorithms Siguang Zhao 1,2, Mingnian Wang 1,2, Wenhao Yi 1,2, Di Yang 1,2 and Jianjun Tong 1,2,* 1 Key Laboratory of Transportation Tunnel Engineering of Ministry of Education, Southwest Jiaotong University, Chengdu 610031, China; 2019310044@my.swjtu.edu.cn (S.Z.); wangmingnian@swjtu.edu.cn (M.W.); 1256635900@my.swjtu.edu.cn (W.Y.); 2020210212@my.swjtu.edu.cn (D.Y.) 2 School of Civil Engineering, Southwest Jiaotong University, Chengdu 610031, China * Correspondence: jjtong@swjtu.edu.cn; Tel.: +86-135-5031-1331 Abstract: The quality evaluation of the surrounding rock is the cornerstone of tunnel design and construction. Previous studies have confirmed the existence of a relationship between drilling parameters and the quality of surrounding rock. The application of drilling parameters to the intelligent classification of surrounding rock has the natural advantages of automatic information collection, real-time analysis, and no extra work. In this work, we attempt to establish the intelligent surrounding rock classification model and software system driven by drilling parameters. We collected 912 samples containing four drilling parameters (penetration velocity, hammer pressure, rotation pressure, and feed pressure) and three surrounding rock (grade-III, grade-IV, and grade-V). Based on the python machine learning toolkit (Scikit-learn), 10 types of supervised machine learning algorithms were used to train the intelligent surrounding rock classification model with the model parameter selection technology of grid search cross validation. The results show that the average accuracy is 0.82, which proves the feasibility of this method. Finally, the tunnel surrounding rock intelligent classification system was established based on three models with better comprehensive performance among them. The classification accuracy of the system was 0.87 in the tunnel test section, which indicates that the system has good generalization performance and practical value. Keywords: drill and blast tunnel; machine learning; measure-while-drilling; drilling parameters; intelligent surrounding rock classification model 1. Introduction Unlike aboveground engineering, in tunnel engineering, tunnels are built under- ground. Tunnel design and construction are closely related to the stratum where it is buried. The quality evaluation of the surrounding rock is the cornerstone of tunnel design [1–3] and construction. Surrounding rock classification is a main evaluation method for the surrounding rock quality. It usually collects the information of surrounding rock by one or more means, and finally gives a comprehensive evaluation index based on specific rule. It can reflect the strength characteristics and deformation characteristics of the surrounding rock and stability characteristics of the tunnel face, and can be directly used to guide the tunnel design and construction. The classification of surrounding rock is a common method for the surrounding rock quality evaluation of tunnels in various countries. This method was started in Europe in 1774. After more than two hundred years of development, there have been hundreds of methods employed for this purpose, such as the Q-value method [4], the rock mass rating (RMR) method [5], and the surrounding rock basic quality index (BQ) method [6]. At present, the methods used to classify the rock surrounding tunnels are mainly qualitative, with quantitative verification using laboratory tests requiring extensive profes- sional knowledge and engineering experience. Therefore, there is some randomness in the Appl. Sci."}], "source_title": "Intelligent Classification of Surrounding Rock of Tunnel Based on 10 Machine Learning Algorithms", "source_doi": "10.3390/app12052656", "source_url": "https://www.mdpi.com/2076-3417/12/5/2656/pdf?version=1646383280", "source_license": "cc-by-sa", "zotero_item_key": "UGAERXYU"}
{"id": "de3b4aeafd8e8780", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: How to evaluate performance of prediction methods? Measures and their interpretation in variation effect analysis; DOI: 10.1186/1471-2164-13-s4-s2; license: cc-by."}, {"role": "assistant", "content": "Vihinen BMC Genomics 2012, 13(Suppl 4):S2 http://www.biomedcentral.com/1471-2164/13/S4/S2 Page 9 of 10\n[page 10] This article has been published as part of BMC Genomics Volume 13 Supplement 4, 2012: SNP-SIG 2011: Identification and annotation of SNPs in the context of structure, function and disease. The full contents of the supplement are available online at http://www.biomedcentral.com/ bmcgenomics/supplements/13/S4. Author details 1Institute of Biomedical Technology, FI-33014 University of Tampere, Finland. 2BioMediTech, Tampere, Finland. 3Department of Experimental Medical Science, Lund University, SE-221 84 Lund, Sweden. Competing interests The author declares that they have no competing interests in relation to the SNP-SIG issue article. Published: 18 June 2012 References 1. Thusberg J, Vihinen M: Pathogenic or not? And if so, then how? Studying the effects of missense mutations using bioinformatics methods. Hum Mutat 2009, 30:703-714. 2. Khan S, Vihinen M: Performance of protein stability predictors. Hum Mutat 2010, 31:675-684. 3. Potapov V, Cohen M, Schreiber G: Assessing computational methods for predicting protein stability upon mutation: good on average but not in the details. Protein Eng Des Sel 2009, 22:553-560. 4. Desmet F, Hamroun G, Collod-Beroud G, Claustres M, Beroud C: Res. Adv. in Nucleic Acid Research. Global Research Network; 2010. 5. Thusberg J, Olatubosun A, Vihinen M: Performance of mutation pathogenicity prediction methods on missense variants. Hum Mutat 2011, 32:358-368. 6. Moult J, Fidelis K, Kryshtafovych A, Tramontano A: Critical assessment of methods of protein structure prediction (CASP)–round IX. Proteins 2011, 79(Suppl 10):1-5. 7. Rodrigues AP, Grant BJ, Godzik A, Friedberg I: The 2006 automated function prediction meeting. BMC Bioinformatics 2007, 8(Suppl 4):S1-4. 8. Wodak SJ: From the Mediterranean coast to the shores of Lake Ontario: CAPRI’s premiere on the American continent. Proteins 2007, 69:697-698. 9. McClure MA, Vasi TK, Fitch WM: Comparative analysis of multiple protein- sequence alignment methods. Mol Biol Evol 1994, 11:571-592. 10. Thompson JD, Plewniak F, Poch O: BAliBASE: a benchmark alignment database for the evaluation of multiple alignment programs. Bioinformatics 1999, 15:87-88. 11. Mizuguchi K, Deane CM, Blundell TL, Overington JP: HOMSTRAD: a database of protein structure alignments for homologous families. Protein Sci 1998, 7:2469-2471. 12. Raghava GP, Searle SM, Audley PC, Barber JD, Barton GJ: OXBench: a benchmark for evaluation of protein multiple sequence alignment accuracy. BMC Bioinformatics 2003, 4:47. 13. Edgar RC: MUSCLE: multiple sequence alignment with high accuracy and high throughput. Nucleic Acids Res 2004, 32:1792-1797. 14. Van Walle I, Lasters I, Wyns L: SABmark–a benchmark for sequence alignment that covers the entire known fold space. Bioinformatics 2005, 21:1267-1268. 15. Orengo CA, Michie AD, Jones S, Jones DT, Swindells MB, Thornton JM: CATH–a hierarchic classification of protein domain structures. Structure 1997, 5:1093-1108. 16. Kolodny R, Koehl P, Levitt M: Comprehensive evaluation of protein structure alignment methods: scoring by geometric measures. J Mol Biol 2005, 346:1173-1188. 17. Lo Conte L, Ailey B, Hubbard TJ, Brenner SE, Murzin AG, Chothia C: SCOP: a structural classification of proteins database. Nucleic Acids Res 2000, 28:257-259. 18. Sonego P, Pacurar M, Dhir S, Kertesz-Farkas A, Kocsor A, Gaspari Z, Leunissen JA, Pongor S: A protein classification benchmark collection for machine learning. Nucleic Acids Res 2007, 35:D232-236. 19. Hwang H, Vreven T, Janin J, Weng Z: Protein-protein docking benchmark version 4.0. Proteins 2010, 78:3111-3114. 20. Cope LM, Irizarry RA, Jaffee HA, Wu Z, Speed TP: A benchmark for Affymetrix GeneChip expression measures. Bioinformatics 2004, 20:323-331. 21. Zhu Q, Miecznikowski JC, Halfon MS: Preferred analysis methods for Affymetrix GeneChips. II. An expanded, balanced, wholly-defined spike- in dataset. BMC Bioinformatics 2010, 11:285. 22. Gray J: The Benchmark Handbook for Database and Transaction Systems. Morgan Kaufmann; 1993. 23. Aniba MR, Poch O, Thompson JD: Issues in bioinformatics benchmarking: the case study of multiple sequence alignment. Nucleic Acids Res 2010, 38:7353-7363. 24. Berman HM, Westbrook J, Feng Z, Gilliland G, Bhat TN, Weissig H, Shindyalov IN, Bourne PE: The Protein Data Bank. Nucleic Acids Res 2000, 28:235-242. 25."}], "source_title": "How to evaluate performance of prediction methods? Measures and their interpretation in variation effect analysis", "source_doi": "10.1186/1471-2164-13-s4-s2", "source_url": "https://bmcgenomics.biomedcentral.com/counter/pdf/10.1186/1471-2164-13-S4-S2", "source_license": "cc-by", "zotero_item_key": "LKJJ9WPZ"}
{"id": "ba3d4dd6f9bb4931", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Spectral Angle Mapping and AI Methods Applied in Automatic Identification of Placer Deposit Magnetite Using Multispectral Camera Mounted on UAV; DOI: 10.3390/min12020268; license: cc-by-sa."}, {"role": "assistant", "content": "SAM Analysis Outputs From the three UAV flight elevations, multispectral images were analyzed and the results show that by varying the reference cosine similarity threshold values, a precise mapping of the area can be attained. By altering their threshold values, Shafri et al. [17] too were able to attain better subject separation. It is clear to see that the SAM analysis performed well across all the three UAV flight elevations from visual analysis of Figure 5. When the threshold value is small, the mask too becomes small, resulting in less of the targeted magnetite iron sands being identified. This in turn means none of the non- magnetite pixels become misidentified. However, where the threshold is larger, more magnetite iron sand pixels are identified. As a result, non-magnetite pixels too are likely to be misidentified as magnetite. This means that one has to manually adjust the threshold such that the mask flawlessly covers the magnetite pixels only.\n[page 9] Minerals 2022, 12, 268 9 of 19 Figure 5. SAM analysis at 2 m, 10 m and 20 m UAV drone elevations at different reference spectral thresholds. The red points represent the user-selected reference point known to be magnetite (pre- SAM), whilst the blue overlay represent areas within the threshold limit thought to be magnetite (post SAM cosine similarity). The best threshold cosine similarities are 0.12, 0.13 and 0.17 at 2 m, 10 m and 20 m, respectively. The x and y axis represent the size of each captured multispectral image. The bar on the right of each first image represents the variation in spectral intensities present in each captured image. Reprinted with permission from ref. [1]. 2021 AusIMM [1] (p. 38).\n[page 10] Minerals 2022, 12, 268 10 of 19 Based on Figure 5, the best SAM cosine similarity threshold values from visual inter- pretation were found to be 0.12, 0.13 and 0.17 for 2 m, 10 m and 20 m UAV flight elevations, respectively. From these results, it can hence be said that distance from the sensor to the sub- ject affects the nature of the SAM analysis masks, given that SAM values required alteration for each flight elevation. In summary, at lower UAV drone altitudes, a lower threshold was necessary to identify magnetite iron sands, whereas the opposite was correct at higher UAV drone altitudes. Parallax error was accounted for as previously described in Section 2.2 due to the multispectral sensors being located on different positions on the multispectral camera (Figure 2). This in turn results in a perfect overlay of all six multispectral images at each UAV flight elevation. Additionally, with larger threshold values, more noise is introduced [16]. This noise is usually caused by mini-ripples and shadows. The SAM analysis is tricked by these ripples into assuming there is resemblance in image pixel characteristics between the actual reference subject and those expressed by ripples and shadows [16]. Even though a visual interpretation of the SAM mask maps seems somewhat instinctive, it can still be argued that this method provides valuable information pertaining to the location of a specific subject such as the magnetite iron sands. There is a plethora of applications within the mining industry where application of this UAV drone coupled with SAM analysis has the potential to be employed. Examples include resource exploration, resource mapping, resource estimation, mine site monitoring, and ore processing amongst others, where quick and accurate separation of rocks and/or minerals is frequently undertaken. 3.3. Application of AI Methods in Magnetite Spectral Classification Post SAM AI algorithms are a trusted technique by which human error can be eliminated in an objective analysis of data [15], this in turn removes subjectivity pertaining to the visual interpretation of SAM segmentation maps. For each segmentation map, pixels within the 0.12, 0.13 and 0.17 cosine similarity thresholds for the 2 m, 10 m and 20 m flight elevations, respectively, are labelled as magnetite. The rest of the pixel data that are outside the thresholds for each flight elevation are labelled as non-magnetite. Figures 6 and 7 are representations of how the segmentation maps were extracted via SAM (shown as binarized images for easy visualization), hence creating the labelled segmentation maps across six spectral bands."}], "source_title": "Spectral Angle Mapping and AI Methods Applied in Automatic Identification of Placer Deposit Magnetite Using Multispectral Camera Mounted on UAV", "source_doi": "10.3390/min12020268", "source_url": "https://www.mdpi.com/2075-163X/12/2/268/pdf?version=1645347690", "source_license": "cc-by-sa", "zotero_item_key": "VPIE7856"}
{"id": "819d279190fe3e9e", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Reinforcement learning based process optimization and strategy development in conventional tunneling; DOI: 10.1016/j.autcon.2021.103701; license: cc-by."}, {"role": "assistant", "content": "Deep Q-learning is a deep reinforcement learning technique that extends the capabilities of clas­ sical Q-learning [34,35] by replacing the value iteration in the Q-table with the function approximator of deep artificial neural networks (ANN). Although applications of ANNs for geotechnical purposes are still often seen as complementary to conventional computational models [36,37], in this case the use of ANNs allows for applications in complex and continuous states spaces while classical Q-learning is confined to discrete states. DQN algorithms are off-policy, model free RL techniques following the Bellman equation (Eq. (1) after [14]) where the optimal action-value function Q*(s,a) is based on a state s and after having taken an action a. The best action is chosen by maximizing the expected value of r + γQ*(s′,a′) where r is the reward, γ is the discount factor that determines how important the future reward is to the algorithm (see also Table 2), s′ is the state at the next time step and a′ are all possible actions [14]. Q*(s, a) = Es′ [ r + γmax a′ Q*(s ′, a ′) |s, a ] (1) [14] have shown that the DQN algorithm can be used for a wide range of different RL problems and albeit there are improvements to deep Q- learning we chose the original implementation as we see it as well suited to establish a baseline for further developments in geotechnical RL. Our implementation is based on the DQN after [14] and the custom DQN implementation from [38]. The main deviations of the network archi­ tecture in comparison to [14,38] are due to the shape of the input and output data with the input being an 2 × 2100 × 2 array and the output a vector of length 8. The number of hidden layers and the decreasing kernel size from the top to the bottom convolutional layers is in accor­ dance with [14] as is the size of the kernel’s stride which is half the size of the kernel itself. While [14] used 32, 64 and 64 filters for each of the three hidden layers respectively, we used 32, 64 and 32 filters for these layers, as we observed that the agent’s performance did not suffer from this decrease, while the computational speed increased. In accordance with these authors we used rectified linear units (ReLU) [39] activation functions. ReLU activation functions have been widely adopted for ANNs within the past decade as they have shown to achieve a better performance than previously used activation functions like the sigmoid (see [40]). As given in [14] the general DQN’s ANN architecture is that of a deep convolutional neural network [41] whose hierarchical struc­ ture mimics the effect of receptive fields and is inspired by [42]. Like [14] we did not perform systematic hyperparameter tuning by random search or similar techniques (see e.g. [43]) due to the big computational effort of the simulation. Hyperparameters were thus optimized manually throughout the course of the development of the RL-simulation. We nevertheless point out that the given DQN architecture as well as the used hyperparameters still have room for improvement (see outlook in Section 7). Table 2 lists all hyperparameters for our DQN implementation. As we use the same terminology as [14] the reader is referred to this paper for more information on the individual parameters. We implemented the DQN using the tensorflow [44] based Python library Keras [45]. Training was done on a NVIDIA GeForce RTX 2080 Ti. From input to output, the agent’s architecture goes as follows and a graphical representation is given in Fig. 5 (with adaptions of architec­ ture from [14] as described in the section ahead): ▪ The input consists of an array with the shape 2 × 2100 × 2 (see Section 3.3.1). ▪ One convolutional layer, with 32 filters, a kernel size of 1 × 16 and a stride of 1 and 8 applying a ReLU activation function [39] ▪ One convolutional layer, with 64 filters, a kernel size of 1 × 8 and a stride of 1 and 4 applying a ReLU activation function ▪ One convolutional layer, with 32 filters, a kernel size of 1 × 4 and a stride of 1 and 2 applying a ReLU activation function ▪ One fully connected layer with 256 neurons applying a ReLU activation function ▪ One fully connected layer with 8 neurons (one per action) as the output layer which applies a linear activation Table 1 The mechanical parameters and permeability of the two ground types, where gt1 represents weak rock and gt2 stronger rock."}], "source_title": "Reinforcement learning based process optimization and strategy development in conventional tunneling", "source_doi": "10.1016/j.autcon.2021.103701", "source_url": "https://www.sciencedirect.com/science/article/pii/S0926580521001527", "source_license": "cc-by", "zotero_item_key": "N5Q33JID"}
{"id": "859d29de8c0c177f", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Automated Excavator Based on Reinforcement Learning and Multibody System Dynamics; DOI: 10.1109/access.2020.3040246; license: cc-by-sa."}, {"role": "assistant", "content": "2016, doi: 10.1016/j.autcon.2016.05.009. [3] T. Hester, TEXPLORE: Temporal Difference Reinforcement Learning for Robots and Time-Constrained Domains, vol. 503. Cham, Switzerland: Springer, 2013, p. 164, doi: 10.1007/978-3-319-01168-4. [4] R. Sutton, F. Bach and A. Barto, Reinforcement Learning, 2nd ed. Cam- bridge, MA, USA: MIT Press, 2018, pp. 1–4. [5] X. Shi, P. J. A. Lever, and F.-Y. Wang, ‘‘Experimental robotic excavation with fuzzy logic and neural networks,’’ in Proc. IEEE Int. Conf. Robot. Autom., Apr. 1996, pp. 957–962, doi: 10.1109/robot.1996.503896. [6] A. Stentz, J. Bares, S. Singh, and P. Rowe, ‘‘Robotic excavator for autonomous truck loading,’’ Auto. Robots, vol. 7, no. 2, pp. 175–186, 1999, doi: 10.1023/A:1008914201877. [7] H. Shao, H. Yamamoto, Y. Sakaida, T. Yamaguchi, Y. Yanagisawa, and A. Nozue, ‘‘Automatic excavation planning of hydraulic excavator,’’ in Intelligent Robotics and Applications (Lecture Notes in Computer Sci- ence), vol. 5315. Berlin, Germany: Springer, 2008, pp. 1201–1211, doi: 10.1007/978-3-540-88518-4_128. [8] D. Schmidt, M. Proetzsch, and K. Berns, ‘‘Simulation and control of an autonomous bucket excavator for landscaping tasks,’’ in Proc. IEEE Int. Conf. Robot. Autom., May 2010, pp. 5108–5113, doi: 10.1109/ ROBOT.2010.5509546. [9] B. J. Hodel, ‘‘Learning to operate an excavator via policy optimization,’’ Procedia Comput. Sci., vol. 140, pp. 376–382, Jan. 2018, doi: 10.1016/j. procs.2018.10.301. [10] S. Dadhich, U. Bodin, F. Sandin, and U. Andersson, ‘‘Machine learn- ing approach to automatic bucket loading,’’ in Proc. 24th Medit. Conf. Control Autom. (MED), Jun. 2016, pp. 1260–1265, doi: 10.1109/MED. 2016.7535925. [11] Q. Zhu and Q.-F. Wang, ‘‘Real-time energy management controller design for a hybrid excavator using reinforcement learning,’’ J. Zhejiang Univ.-Sci. A, vol. 18, no. 11, pp. 855–870, Nov. 2017, doi: 10.1631/ jzus.A1600650. [12] J. Tobin, R. Fong, A. Ray, J. Schneider, W. Zaremba, and P. Abbeel, ‘‘Domain randomization for transferring deep neural networks from simu- lation to the real world,’’ in Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst. (IROS), 2017. Accessed: Nov. 25, 2020, doi: 10.1109/iros.2017.8202133. [13] J. Xu, T. Du, M. Foshey, B. Li, B. Zhu, A. Schulz, and W. Matusik, ‘‘Learning to fly?: Computational controller design for hybrid UAVs with reinforcement learning learning to fly?: Computational controller design for hybrid UAVs with reinforcement learning,’’ ACM Trans. Graph., vol. 38, no. 4, pp. 1–12, Jul. 2019, doi: 10.1145/3306346.3322940. [14] Z. Xie, P. Clary, J. Dao, P. Morais, J. Hurst, and M. Van De Panne, ‘‘Iterative reinforcement learning based design of dynamic locomo- tion skills for cassie,’’ 2019, arXiv:1903.09537. [Online]. Available: https://arxiv.org/abs/1903.09537 [15] P. Hämäläinen, A. Babadi, X. Ma, and J. Lehtinen, ‘‘PPO-CMA: Prox- imal policy optimization with covariance matrix adaptation,’’ 2018, arXiv:1810.02541. [Online]. Available: http://arxiv.org/abs/1810.02541 [16] J. de García Jalón, E. Álvarez, F. A. de Ribera, I. Rodríguez, and F. J. Funes, ‘‘A fast and simple semi-recursive formulation for multi-rigid- body systems,’’ in Advances in Computational Multibody Systems. Dor- drecht, The Netherlands: Springer, 2005, pp. 1–23. VOLUME 8, 2020 214005\n[page 9] I. Kurinov et al.: Automated Excavator Based on RL and Multibody System Dynamics [17] A. Avello, J. M. Jiménez, E. Bayo, and J. G. de Jalón, ‘‘A simple and highly parallelizable method for real-time dynamic simulation based on velocity transformations,’’ Comput. Methods Appl. Mech. Eng., vol. 107, no. 3, pp. 313–339, Aug. 1993, doi: 10.1016/0045-7825(93)90072-6. [18] J. G. de Jalón and E. Bayo, Kinematic and Dynamic Simulation of Multi- body Systems: The Real-Time Challenge. New York, NY, USA: Springer, 1993. [19] Y. Pan, W. Dai, Y. Xiong, S. Xiang, and A. Mikkola, ‘‘Tree-topology- oriented modeling for the real-time simulation of sedan vehicle dynam- ics using independent coordinates and the rod-removal technique,’’ Mechanism Mach. Theory, vol. 143, Jan. 2020, Art. no. 103626, doi: 10.1016/j.mechmachtheory.2019.103626. [20] J. Watton, Fluid Power Systems?: Modeling, Simulation, Analog and Microcomputer Control. Upper Saddle River, NJ, USA: Prentice-Hall, 1989. [21] H. M. Handroos and M. J. Vilenius, ‘‘Flexible semi-empirical models for hydraulic flow control valves,’’ J. Mech. Design, vol. 113, no. 3, pp. 232–238, Sep. 1991, doi: 10.1115/1.2912774. [22] M. Pla-Castells, I."}], "source_title": "Automated Excavator Based on Reinforcement Learning and Multibody System Dynamics", "source_doi": "10.1109/access.2020.3040246", "source_url": "https://ieeexplore.ieee.org/ielx7/6287639/6514899/09268069.pdf", "source_license": "cc-by-sa", "zotero_item_key": "7F2IQNFD"}
{"id": "adbef2f2dc8dc16a", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Detection of Ionospheric Scintillation Based on XGBoost Model Improved by SMOTE-ENN Technique; DOI: 10.3390/rs13132577; license: cc-by-sa."}, {"role": "assistant", "content": "Resampling Technique Raw Data Random Oversampling Random Undersampling SMOTE ENN SMOTE-ENN Precision (%) 0 99.38 99.51 99.64 99.77 99.47 99.71 1 82.45 83.07 85.00 85.05 86.67 87.96 2 96.02 95.14 94.83 95.19 95.77 95.00 Recall (%) 0 87.91 87.55 89.43 89.31 92.13 92.16 1 97.89 97.64 97.65 97.93 97.85 97.77 2 86.55 90.15 90.75 91.17 87.47 91.60 F1 score (%) 0 93.29 93.15 94.26 94.25 95.66 95.79 1 89.51 89.77 90.89 91.04 91.93 92.61 2 91.01 92.58 92.75 93.14 91.43 93.27 Accuracy (%) 91.43 91.69 92.69 92.80 93.58 94.17 Based on the SMOTE-ENN resampling technique, Table 5 compares the experimental data for five detection algorithms. The decision tree and random forest, each of which presented excellent performance in the binary classification for ionospheric scintillation detection in recent research [16], were tested for comparison. Considering the good per- formance of the SVM in similar detection tasks [19–23], tests for such a method were performed and analyzed. The CNN model was also employed in the comparison due to its effectiveness in a wide range of classification problems [25–28]. However, the SMOTE-ENN resampling technique is meaningless in regard to improving CNN performance, as its space structure becomes disordered after the resampling process. Thus, the CNN test was carried out on raw data only, while the same tests on the SVM accord with those on the other three methods. Due to the mentioned problem of the CNN and excessive computational loads of the SVM, these two methods were not studied further. As shown in\n[page 14] Remote Sens. 2021, 13, 2577 14 of 22 Table 5, apart from the overall accuracy, the recall and F1 score for class 2 were recorded to illustrate the performance in detecting strong scintillation events. In addition, considering the unavailability of the resampling technique for the CNN, larger computational loads and running time for the SVM, and the overall better performance with XGBoost, further tests with both methods were not carried out. Further details can be found in the Discussion. Table 5. Comparison of performance of different detection algorithms trained with raw dataset and SMOTE-ENN resampled dataset (apart from CNN algorithm), based on testing results. Algorithm Accuracy (%) Recall for Class 2 (%) F1 Score for Class 2 (%) Raw Data SMOTE- ENN Improvement Ratio Raw Data SMOTE- ENN Improvement Ratio Raw Data SMOTE- ENN Improvement Ratio CNN 89.86 - - 73.63 - - 82.15 - - SVM 93.39 93.52 0.13 90.49 95.06 4.57 89.46 89.07 -0.39 Decision tree 85.39 87.36 1.97 86.73 90.41 3.68 88.94 90.98 2.04 Random forest 86.95 88.92 1.97 85.44 89.97 4.53 90.43 92.91 2.48 XGBoost 91.43 94.17 2.74 86.55 91.60 5.05 91.04 93.27 2.23 When training on the raw data and resampled data, the accuracy of XGBoost was about 4% and 7% better than that of the decision tree and random forest, as shown in the left panel of Figure 8. The middle and right panels in Figure 8 also show higher recall and F1 scores for the XGBoost method, implying the superiority of XGBoost improved by the SMOTE-ENN technique. 85 87 89 91 93 95 Raw data SMOTE-ENN 2.74% growth (a) 86 88 90 92 Raw data SMOTE-ENN (b) 5.05% growth 89 90 91 92 93 Raw data SMOTE-ENN 2.23% growth (c) Figure 8. Comparison of detection algorithms in terms of overall accuracy, recall value for class 2 and F1 scores for class 2, respectively. From left to right: (a) the overall accuracy; (b) the recall value for class 2; (c) the F1 score for class 2. The blue polyline marked by triangles refers to the testing results based on raw training data, while the orange polyline marked by asterisks represents the testing results based on SMOTE-ENN resampled training data. The 100,000 training samples were selected in the overall training dataset listed in Table 1, while the testing dataset was that mentioned in Table 1. 4.4. Analysis on Imbalanced Training Datasets To explore the detection performance of the XGBoost algorithm improved by the SMOTE-ENN resampling technique for strong scintillation events, a series of comparison experiments were performed based on different degrees of imbalance in the training and testing datasets, compared with the decision tree and random forest algorithm. From the training dataset with 349,559 samples, seven subsets of data were extracted. For each subset, 100,000 randomly selected samples were included, with the ratio of classes 0, 1 and 2 ranging from 1:1:1 to 7:7:1, respectively."}], "source_title": "Detection of Ionospheric Scintillation Based on XGBoost Model Improved by SMOTE-ENN Technique", "source_doi": "10.3390/rs13132577", "source_url": "https://www.mdpi.com/2072-4292/13/13/2577/pdf?version=1625212202", "source_license": "cc-by-sa", "zotero_item_key": "4YEIWMK3"}
{"id": "f5c559497c041d10", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Reinforcement learning based process optimization and strategy development in conventional tunneling; DOI: 10.1016/j.autcon.2021.103701; license: cc-by."}, {"role": "assistant", "content": "A solution would be to take a version of the agent that was saved before instability occurred and continue training with that, but we left Fig. 7. Different training paths for five exemplary agents over 120,000 episodes. The first row shows cumulative rewards per episode; the second row the average number of blasts/moves that were required to complete each episode; the third row the average number of face instabilities per episode; the fourth row shows the average loss per episode. Transparent colors in the background show the raw data records and the solid lines in the foreground a 500-episode sliding window average. Fig. 8. Histogram for the performance of an agent who plays 10,000 episodes with completely random moves (i.e. ε = 1). G.H. Erharter et al.\n[page 9] Automation in Construction 127 (2021) 103701 9 the agents 2 and 5 as they are for explanatory purposes. Before discussing the performance of the agents, we first establish a baseline for the minimum performance that an agent must reach by letting an agent play with completely random moves (i.e. ε = 1) for 10,000 episodes. In these 10,000 random episodes the agent reached a ▪ minimum/maximum/median reward of −544/255/-13 points, ▪ minimum/maximum/median number of 0/27/8 instabilities per episode, ▪ minimum/maximum/median number of 124/197/145 blasts/ moves per episode. Histograms for the 10,000 random episodes are given in Fig. 8. The maximum rewards, minimum number of face instabilities and minimum number of blasts/moves that the five agents of Fig. 7 needed are given in Table 5. We computed these numbers based on sliding window averages of 500 episodes as to avoid individual episodes that performed extraordinarily well. Albeit substantial differences can be seen in these statistics, it can generally be observed that all maximum rewards of the agents are above the maximum rewards of the randomly played episodes and therefore training was generally a success. In these five agents, two groups can be observed, where the agents 2 and 3 reached their maximum performance between episodes 40,000 and 60,000 and the agents 1, 4 and 5 reached their maximum perfor­ mance between episodes 80,000 and 120,000. Although the first group has reached their peak performance sooner, the achieved reward is generally lower than that of the second group (see description of training paths above). We conclude this section by presenting the strategy that the best performing agent 4 has found after 119,000 episodes – at its peak per­ formance (in terms of highest reward and low instabilities). We tested the saved agent’s checkpoint for 10,000 episodes with a fixed ε of 0.05 which corresponds to the ε at that stage of training and is in accordance to [14] who recommend an ε > 0 also for testing, as this helps the agent to deal with unexpected situations. The histograms of Fig. 9 show the same test statistics as given for the random moves in Fig. 8. The strategy that the agent adopted is focused on long advance lengths without face support, in alternation with long advance lengths with face support. By doing so the agent avoids face instabilities, while also minimizing the required support (see Table 4 for the respective rewards/penalties). The boxplot of Fig. 10 illustrates this, as it can be observed that the majority of actions is 4 m long advance lengths. Furthermore, the agent focuses on long advance lengths without face support in both excavation types (top heading and bench) which shows that it tries to avoid excessive use of support measures, thus showing a tendency towards economical optimization. It can also be seen that the agent still uses small advance lengths sometimes and, in this case, favors the actions without face support as it has realized that the small advance lengths do not lead to face instabilities in the given conditions. In Fig. 11 an example of one episode for the agent 4 is visualized. The time-distance diagram in the top row of this figure shows that the agent has learned to optimize the excavation process by minimizing changes between top heading and bench excavation which would be associated with unwanted delays in “real life” tunneling (e.g. building and removing of access ramps). As given in Section 2.1, installed support ahead of the current face always covers 10 m. With the maximum advance length being set to 4 m it would be unnecessary and uneco­ nomical to install face support in two consecutive rounds. In the second row of Fig."}], "source_title": "Reinforcement learning based process optimization and strategy development in conventional tunneling", "source_doi": "10.1016/j.autcon.2021.103701", "source_url": "https://www.sciencedirect.com/science/article/pii/S0926580521001527", "source_license": "cc-by", "zotero_item_key": "N5Q33JID"}
{"id": "30e1cfe4710514e5", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Intelligent Classification of Surrounding Rock of Tunnel Based on 10 Machine Learning Algorithms; DOI: 10.3390/app12052656; license: cc-by-sa."}, {"role": "assistant", "content": "The tunnel is 5498 m in length with a maximum buried depth of 320 m. The main lithologies of the surrounding rock are limestone and shale, and the grades of the surrounding rocks are III (1493 m), IV (1050 m), and V (720 m). The method of construction of this tunnel is the full-section method with large ma- chinery, i.e., the tunnel face (150 m2) is excavated and formed by blasting once, and the primary support is closed once. With this method, the circular footage is 2–4.8 m, and the monthly progress is 60–150 m in this tunnel. The tunnel position is shown in Figure 12. This test section is DK450 + 834~DK451 + 126, which is 292 m in total length. The excavation revealed that the surrounding rock was grade-III; the lithology was bluish grey, a massive structure, and hard limestone rock. The stability of the tunnel face was good without developed joints, fractures, or groundwater. The test section of this tunnel is shown in Figure 13, and the tunnel face is shown in Figure 14.\n[page 16] Appl. Sci. 2022, 12, 2656 16 of 20 Figure 12. Interface of the position of the Gaojiaping Tunnel of the Zhengzhou–Wanzhou high- speed railway. DK451+126 Tunnel Entrance mileage:DK449+037 DK450+834 DK451+000 Limestone(T1d) Exit mileage:DK454+535 Wanzhou Zhenzhou 650 450 250 500 300 H（m） 550 350 600 400 200 Figure 13. Test section of Gaojiaping Tunnel.\n[page 17] Appl. Sci. 2022, 12, 2656 17 of 20 Figure 14. Tunnel face of this test section of the Gaojiaping Tunnel. 3.3. Intelligent Surrounding Rock Classification in Test Tunnel Among the test section, drilling parameter information from 30 tunnel faces was col- lected by the intelligent surrounding rock classification system. The intelligent classification results of the surrounding rock on each tunnel face are shown in Table 8. Table 8. Intelligent classification of surrounding rock on each tunnel face. Number Mileage Predicted Results Actual Results Right/Error 1 DK450+834 III III Right 2 DK450+837 III III Right 3 DK450+840 III III Right 4 DK450+916 III III Right 5 DK450+919 III III Right 6 DK450+922 III III Right 7 DK450+926 III III Right 8 DK451+014 III III Right 9 DK451+017 III III Right 10 DK451+021 III III Right 11 DK451+024 III III Right 12 DK451+027 III III Right 13 DK451+030 III III Right 14 DK451+034 III III Right 15 DK451+037 III III Right 16 DK451+042 III III Right 17 DK451+046 III III Right 18 DK451+051 III III Right 19 DK451+055 III III Right 20 DK451+060 III III Right 21 DK451+064 III III Right 22 DK451+100 III III Right 23 DK451+103 III III Right 24 DK451+106 III III Right 25 DK451+109 III III Right 26 DK451+113 IV III Error 27 DK451+116 IV III Error 28 DK451+119 IV III Error 29 DK451+123 IV III Error 30 DK451+126 III III Right Accuracy 86.7%\n[page 18] Appl. Sci. 2022, 12, 2656 18 of 20 In Table 8, the classification accuracy of the surrounding rock intelligent classification system on-site is 86.7%, which indicates that the system has good generalization. However, due to the single lithology and surrounding rocks (limestone, grade-III) in the field test section, its generalization performance must be further verified. 4. Conclusions Based on 912 drilling parameters of the Zhengzhou–Wanzhou high-speed railway tunnel project, 10 intelligent surrounding rock classification models were established using multiple machine learning algorithms. With three models (SVM, RT, and ET), an intelligent surrounding rock classification system was established and verified by the field test section. The main conclusions of this study are: 1. All the absolute values of the correlation coefficients between the four drilling pa- rameters (Vp, Ph, Pr, and Pf) and the surrounding rock grades (III, IV, and V) are above 0.5. Thus, there is a correlation between the drilling parameters and the surrounding rock grade, so the method of surrounding rock classification by drilling parameters is feasible. 2. The 10 models based on supervised machine learning algorithms all have good performance. The average accuracy of them is 0.82. In particular, the average recalls of grades-III and -V were greater than 0.90. Thus, these machine learning models established by drilling parameters are feasible and reliable in the intelligent classification of surrounding rocks. 3. By comprehensively considering precision, recall, F1 score, and accuracy, we observe that three types of models containing SVM, RT, and ET have better performance among 10 machine learning algorithm models."}], "source_title": "Intelligent Classification of Surrounding Rock of Tunnel Based on 10 Machine Learning Algorithms", "source_doi": "10.3390/app12052656", "source_url": "https://www.mdpi.com/2076-3417/12/5/2656/pdf?version=1646383280", "source_license": "cc-by-sa", "zotero_item_key": "UGAERXYU"}
{"id": "608d5673c8bcf667", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Wheel Loader Scooping Controller Using Deep Reinforcement Learning; DOI: 10.1109/access.2021.3056625; license: cc-by-sa."}, {"role": "assistant", "content": "[page 1] Received January 12, 2021, accepted January 20, 2021, date of publication February 2, 2021, date of current version February 10, 2021. Digital Object Identifier 10.1109/ACCESS.2021.3056625 Wheel Loader Scooping Controller Using Deep Reinforcement Learning OSHER AZULAY 1 AND AMIR SHAPIRO 2, (Member, IEEE) 1School of Mechanical Engineering, Tel Aviv University, Tel Aviv 69978, Israel 2Department of Mechanical Engineering, Ben-Gurion University of the Negev, Be’er Sheva 84105, Israel Corresponding author: Osher Azulay (osherazulay@mail.tau.ac.il) This work was supported by the Israeli Ministry of Defense through Project ROBIL under Grant 87726311. ABSTRACT This article presents a deep reinforcement learning-based controller for an unmanned ground vehicle with a custom-built scooping mechanism. The robot’s aim is to autonomously perform earth scooping cycles with three degrees of freedom: lift, tilt and the robot’s velocity. While the majority of previous studies on automated scooping processes are based on data recorded by expert operators, we present a method to autonomously control a wheel loader to perform the scooping cycle using deep reinforcement learning methods without any user-provided demonstrations. The controller’s learning approach is based on the actor- critic, Deep Deterministic Policy Gradient algorithm which we use to map online sensor data as input to continuously update the actuator commands. The training of the scooping policy network is done solely in a simplified simulation environment using a virtual physics engine, which converges to an average of a 65% fill factor from the full bucket capacity and a 5 [sec] average cycle time. We illustrate the performance of the trained policy in simulations and in real-world experiments with 3 different inclination angles of the earth. An additional scooping experiment compared the performance of our controller to remote manual human control. Overall, the deep reinforcement learning-based controller exhibited good performance in terms of both achieved visually bucket fill with varying scooped earth weights of 4.1 −7.2[kg], and a 5.1 −7.1[sec] cycle time. The experimental results confirm the ability of our planner to fill bucket as required, indicating that our controller can be used for excavation purposes. INDEX TERMS Robotics in construction, machine learning, agricultural automation. I. INTRODUCTION Earthmoving systems are currently used in a variety of industries, but especially in the construction and agricul- tural domains. They have numerous advantages including economic efficiency, safety, and availability. Earthmoving machinery typically refers to heavy-duty vehicles designed for construction operations that involve earthworks. While earthmoving machinery continues to develop, most of the excavation cycle is still controlled by a human, either directly or via teleoperation [1]. Designing control methods for such tasks is a long-standing research goal, which has attracted considerable interest and generated a number of survey papers [2], [3]. The general approaches to autonomous excavation exploit the machine dynamics and try to follow a defined trajectory [4], [5], use compliance force control [6], [7] or employ a behavior-based approach for motion The associate editor coordinating the review of this manuscript and approving it for publication was Yangmin Li . FIGURE 1. A deep reinforcement learning controller was trained in simulation to accomplish the scooping cycle using 12-dimensional sensors inputs into three actuator commands. control [8], [9]. Since some tasks are more complex than oth- ers, they often require extensive engineering experience and tedious manual tuning beyond the control algorithm itself. One of the key challenges in deploying automated earth moving machines relates to the analysis of the soil-tool interaction, due to the unpredictable nature of the soil [6]. VOLUME 9, 2021 This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/ 24145\n[page 2] O. Azulay, A. Shapiro: Wheel Loader Scooping Controller Using Deep Reinforcement Learning Most approaches require accurate models of the machine which makes them liable to modeling errors, wear and tear, and changing conditions [10]."}], "source_title": "Wheel Loader Scooping Controller Using Deep Reinforcement Learning", "source_doi": "10.1109/access.2021.3056625", "source_url": "https://ieeexplore.ieee.org/ielx7/6287639/9312710/09344588.pdf", "source_license": "cc-by-sa", "zotero_item_key": "IDZBIHT7"}
{"id": "4ec1c6be35743a0c", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Wheel Loader Scooping Controller Using Deep Reinforcement Learning; DOI: 10.1109/access.2021.3056625; license: cc-by-sa."}, {"role": "assistant", "content": "Wang, ‘‘Intelligent excavator control system for lunar mining system,’’ J. Aerosp. Eng., vol. 8, no. 1, pp. 16–24, Jan. 1995. [19] A. A. Dobson, J. A. Marshall, and J. Larsson, ‘‘Admittance control for robotic loading: Underground field trials with an LHD,’’ in Field and Service Robotics. Toronto, ON, Canada, 2015. [20] D. Jud, G. Hottiger, P. Leemann, and M. Hutter, ‘‘Planning and control for autonomous excavation,’’ IEEE Robot. Autom. Lett., vol. 2, no. 4, pp. 2151–2158, Oct. 2017. [21] G. J. Maeda, I. R. Manchester, and D. C. Rye, ‘‘Combined ILC and disturbance observer for the rejection of near-repetitive disturbances, with application to excavation,’’ IEEE Trans. Control Syst. Technol., vol. 23, no. 5, pp. 1754–1769, Sep. 2015. VOLUME 9, 2021 24153\n[page 10] O. Azulay, A. Shapiro: Wheel Loader Scooping Controller Using Deep Reinforcement Learning [22] S. Singh and H. Cannon, ‘‘Multi-resolution planning for earthmoving,’’ in Proc. IEEE Int. Conf. Robot. Automat., May 1998, pp. 121–126. [23] O. Kanai, H. Osumi, S. Sarata, and M. Kurisu, ‘‘Autonomous scooping of a rock pile by a wheel loader using disturbance observer,’’ in Proc. 23rd Int. Symp. Autom. Robot. Construct., Oct. 2006, pp. 3–5. [24] S. Dadhich, U. Bodin, F. Sandin, and U. Andersson, ‘‘Machine learning approach to automatic bucket loading,’’ in Proc. 24th Medit. Conf. Control Autom. (MED), Jun. 2016, pp. 1260–1265. [25] E. Halbach, J. Kamarainen, and R. Ghabcheloo, ‘‘Neural network pile loading controller trained by demonstration,’’ in Proc. Int. Conf. Robot. Autom. (ICRA), May 2019, pp. 980–986. [26] R. Jonschkowski and O. Brock, ‘‘Learning state representations with robotic priors,’’ Auto. Robots, vol. 39, no. 3, pp. 407–428, Oct. 2015. [27] T. Lesort, N. Díaz-Rodríguez, J.-F. Goudou, and D. Filliat, ‘‘State rep- resentation learning for control: An overview,’’ Neural Netw., vol. 108, pp. 379–392, Dec. 2018. OSHER AZULAY received the B.Sc. and M.Sc. degrees (Hons.) from the Department of Mechan- ical Engineering, Ben-Gurion University of the Negev, Israel, in 2019 and 2020, respectively. He is currently pursuing the Ph.D. degree with the School of Mechanical Engineering, Tel Aviv University. His current research interests include robotic manipulators and reinforcement learn- ing. He had been awarded numerous prizes for academic and research excellence, including certificates of achievement 2017–2018 and 2018–2019 and an Excellence Scholarship from the Mechanical Engineering Department, Ben-Gurion Uni- versity of the Negev. AMIR SHAPIRO (Member, IEEE) received the B.Sc., M.Sc., and Ph.D. degrees in mechanical engineering from the Technion, Israel Institute of Technology, Haifa, in 1997, 2000, and 2004, respectively. From 2005 to 2006, he was a Postdoc- toral Fellow with the Robotics Institute, Carnegie Mellon University, Pittsburgh, PA, USA. He is cur- rently an Associate Professor and the Director of the Robotics Laboratory, Department of Mechan- ical Engineering, Ben-Gurion University of the Negev, Be’er Sheva, Israel. His research interests include locomotion of multi-limbed mechanisms in unstructured complex environments, motion planning algorithms for multi-limbed robots, robot grasping-design, control, and stability analysis, climbing robots, snake-like robots, multi-robot on-line motion planning, and agricultural robotics. 24154 VOLUME 9, 2021"}], "source_title": "Wheel Loader Scooping Controller Using Deep Reinforcement Learning", "source_doi": "10.1109/access.2021.3056625", "source_url": "https://ieeexplore.ieee.org/ielx7/6287639/9312710/09344588.pdf", "source_license": "cc-by-sa", "zotero_item_key": "IDZBIHT7"}
{"id": "2d16a0c7da498d36", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Intelligent Classification of Surrounding Rock of Tunnel Based on 10 Machine Learning Algorithms; DOI: 10.3390/app12052656; license: cc-by-sa."}, {"role": "assistant", "content": "Field collection of the drilling parameters. (a) Blasting hole drilling in the tunnel face; (b) the cemented carbide bit. The area of the tunnel face is about 150 m2, containing about 200~300 blasting holes. The value of each drilling parameter for each sample is the average of all the drilling holes in the tunnel face. The typical layout of the blasting hole is shown in Figure 5. The grade of the surrounding rock in this study is specified in the current Code for Design of Railway Tunnel (TB10003-2016). According to the degree of hardness, integrity, groundwater state, crustal stress state, and major weak structural surface, the surrounding rocks are divided into Grades I~VI according to the quality. The approximate correspondence between the surrounding rock grade in this study and the indices of other classification methods, such as the Q method, RMR method, and BQ method, is shown in Table 1 [55]. Table 1. Approximate correspondence between the surrounding rock grade in this study and other indices (Q, RMR, and BQ). Method Surrounding Rock Grade I II III IV V Q value (10,000,40) (40,10) (10,1) (1,0.1) (0.1,0.001) RMR value (100,80) (80,60) (60,40) (40,20) (20,0) BQ value (700,550) (550,450) (450,350) (350,250) (250,0) Note: Grade-VI generally refers to the special geology, such as the powdery fault fracture zone, aeolian sand, and seriously collapsible loess.\n[page 6] Appl. Sci. 2022, 12, 2656 6 of 20 Figure 5. The typical layout of the blasting hole (mm). The geological sketch method is usually adopted to determine the grade of the sur- rounding rock in the tunnel, which is analyzed and determined by professional geological engineers. The geological sketch record card of the tunnel face is shown in Table 2. Table 2. Geological sketch record card of the tunnel face. The Geological Sketch Record Card of Tunnel Face Project name: XXX Mileage: XXX Date: XXX Construction unit: XXX No. Item State description 1 Tunnel face and support type Width (m) Area (m2) Excavation method Depth (m) Advanced support Primary support Secondary lining 2 Lithology 3 Weathering degree Slightly Weakly Strongly Totally 4 Degree of hardness (Mpa) Extremely hard (>60) Hard (30~60) Relatively soft (15~30) Soft (5~15) Extremely soft (<5) 5 Number of structural surfaces 1 2 3 >3 6 Average spacing of the structural surface (m) >1 0.4~1 0.2~0.4 0~0.2 7 Degree of structural surface development Not developed Relatively developed Developed Extremely developed 8 Crack width (mm) 0~1 1~3 >3 9 Crack filling None Siliceous Calcium Argillaceous Rock debris Mud clamps rock debris\n[page 7] Appl. Sci. 2022, 12, 2656 7 of 20 Table 2. Cont. The Geological Sketch Record Card of Tunnel Face 10 Degree of structural surface bonding Good Fair Poor Extremely poor 11 Integrity Integrated Relatively integrated Relatively broken Broken Extremely broken 12 Groundwater state None Drip outflow Linear outflow Inrush outflow 13 The major weak structural surface Occurrence Angle with tunnel axis 14 Crustal stress state Low High Extremely high 15 Surrounding rock grade I II III IV V VI Tabulator: XXX Reviewer: XXX By the method discussed above, 912 intelligent surrounding rock classification samples were collected in five tunnels of the Zhengzhou–Wanzhou high-speed railway. These samples cover three surrounding rock grades (grade-III, grade-IV, and grade-V) and five main lithologies (dolomite, sandstone, limestone, shale, and mudstone). More details are shown in Table 3. Table 3. Intelligent surrounding rock classification sample statistics. Surrounding Rock Grade Tunnel Lithology Sample Size Total III Luojiashan Tunnel Dolomite 110 325 Chufeng Tunnel Dolomite 28 Xinhua Tunnel Sandstone 31 Xiangjiawang Tunnel Limestone 156 IV Gaojiaping Tunnel Shale/Sandstone 105 420 Luojiashan Tunnel Dolomite 75 Chufeng Tunnel Dolomite 44 Xinhua Tunnel Sandstone 125 Xiangjiawang Tunnel Limestone 30 Xiangluping Tunnel Mudstone/Sandstone 41 V Gaojiaping Tunnel Shale 84 167 Luojiashan Tunnel Dolomite 62 Chufeng Tunnel Dolomite 4 Xinhua Tunnel Dolomite 15 Xiangjiawan Tunnel Limestone 2 Total 912 2.3. Sample Data Cleaning To analyze the sample data outliers, four box diagrams are shown in Figure 6 according to the drilling parameter class and surrounding rock grade.\n[page 8] Appl. Sci. 2022, 12, 2656 8 of 20 (a) (b) (c) (d) Figure 6."}], "source_title": "Intelligent Classification of Surrounding Rock of Tunnel Based on 10 Machine Learning Algorithms", "source_doi": "10.3390/app12052656", "source_url": "https://www.mdpi.com/2076-3417/12/5/2656/pdf?version=1646383280", "source_license": "cc-by-sa", "zotero_item_key": "UGAERXYU"}
{"id": "588bac15e700cc89", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Wheel Loader Scooping Controller Using Deep Reinforcement Learning; DOI: 10.1109/access.2021.3056625; license: cc-by-sa."}, {"role": "assistant", "content": "In the next phase, the robot changes the lift, tilt, and velocity actions simultaneously to navigate the bucket tip through the earth pile. Lift and tilt angles are increased and velocity decreases while the robot excavates deeper. In the last phase, the robot drives backward away from the pile, while lifting the arm to its maximum height to maintain the soil inside the bucket without loss of the material. B. REAL ENVIRONMENT EVALUATION To investigate the performance and adaptability of our scoop- ing controller in real-life machinery, we conducted several experiments with three different mean inclination angles of the pile. We explored the learned policies trained in simulation in the real world without tuning the network weights. To derive the amount of excavated soil in the bucket, we attached a load sensor to the bucket joint that measured the load on the linear actuator. Under the assumption that acceleration and velocity are negligible during excavation, the scooping dynamics in joint space can be reduced to a simple static equilibrium: τ = G + JT f (8) where f is a vector representing the contact forces, G is a vector containing the gravity terms, τ is a vector used to denote the torque at each joint and J is the Jacobian matrix of the arm. The masses of the links of the arm mechanism are known, as well as the Jacobian matrix, so J and G are obtained through direct computation using the joint posi- tions. f is obtained through the load cell measurements since we can derive a simple quasi-static analysis that describes the moment’s equilibrium at the bucket joint. Although we approximated the accumulated mass using the load cell, as described in Section III, most of the inconsistency between the simulation and real-life occurred during the fill phase, as can be seen in Fig. 10, because the robot basically trans- lated this to additional scooped earth, resulting in an earlier exit phase. FIGURE 10. Example of comparisons between the normalized extracted mass inputs in the simulated (100 episodes) and real-life experiments. The data points are normalized by the maximum measured values of each set. In our experiments, we defined two environment setups to be constant: (1) the position of the pile and (2) the starting position of the robot (1[m] in front of the pile). We set these conditions to keep the problem’s complexity manageable and because these setups are common to various types of local- ization methods and feature detection methods. The bucket scooping trajectories shown in this section were obtained using direct kinematics of the arm mechanism feedback and odometry from the wheel encoders. Computation of the total excavated earth in the experiments was carried out based on the applied torque, where the bucket exits the pile of gravel. To compare the planner’s performance to human operators, the first step in the evaluation consisted of collecting data from manually driven scooping cycles using manual control. The first author collected the data while attempting to man- ually imitate the three phase scoop strategy. Although the goal was to achieve continuous actions, the motions were less skilled than that of an expert operator. In total, 20 trials were conducted, which were less smooth during the first attempts than in the later ones. We evaluated the learned scooping policy over 5 consecutive scooping cycles, where the earth’s inclination angle set to 32 degrees. Fig. 11 and Fig. 12 show the results, where each phase is labelled. In most of the approach phases, the robot prepared to interact with the soil by adjusting the bucket to a horizontal angle and decreasing the lift angle. Each contact with the pile triggered the fill phase, where the robot penetrates the pile with a ‘‘slicing motion’’. During that motion, the applied torque increased due to higher pile resistance. The controller continued the filling phase until a combination of depth and torque thresholds were reached, leading to a low variation exit phase initial- ization. Although we initialized each trial with the same setup, the controller executed different trajectories during the approach phase, due to changes in the soil-tool interaction, VOLUME 9, 2021 24151\n[page 8] O. Azulay, A. Shapiro: Wheel Loader Scooping Controller Using Deep Reinforcement Learning FIGURE 11. Comparisons of 5 consecutive scooping attempts from the pile with an average of 32 degrees of inclination angle. Bucket trajectory (A) applied torque at the bucket joint (B)."}], "source_title": "Wheel Loader Scooping Controller Using Deep Reinforcement Learning", "source_doi": "10.1109/access.2021.3056625", "source_url": "https://ieeexplore.ieee.org/ielx7/6287639/9312710/09344588.pdf", "source_license": "cc-by-sa", "zotero_item_key": "IDZBIHT7"}
{"id": "0fd6b008d49cd39c", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Automated Excavator Based on Reinforcement Learning and Multibody System Dynamics; DOI: 10.1109/access.2020.3040246; license: cc-by-sa."}, {"role": "assistant", "content": "The step method is called for each time step of the simulation. It sends the observation vector, receives the action vector and assigns Algorithm 1 PPO-CMA (Server Side) 1: call function make: 2 for iteration = 1, 2, ...do 3: while iteration simulation budget N not exceeded do 4: call function reset 5: for T timesteps run agent on current policy or until a terminal state 6: call function step 7: end while 8: Train critic network and policy 9: Save (serialize) agent 10: end for Algorithm 2 Step Function in Simulation Environment (Client Side) function step (simulation parameters list): 1: Receive action vector from agent (server) 2: Set actions to simulation model inputs 3: for action 4: Set action value to the hydraulic system input 5: end for 6: for observation parameter 7: Get simulation variables from Solver 8: Compose observation vector 9: Calculate reward 10: Encode reward and observation to a single message 11: end for 12: Send encoded observation vector, reward and done to the agent end function it to the corresponding inputs. The reset method is used to restart the simulation when the simulation budget is exceeded or the termination state has been reached. The Parameters class is used for transfer of simulation-related parameters to the RL algorithm. The State class obtains the observation vector from the simulation and calculates the reward value of the provided action. III. PROBLEM STATEMENT The above procedure was applied to an excavator model with a complicated environment with deformable ground. Fig.4 shows the structure of the model. The model is a system with closed loops and has 4 Degrees of Freedom (DOF). The model consists of a total of nine rigid bodies interconnected by 10 joints. The model is equipped with the hydraulic circuit shown in Fig.5. The excavator model has 4 inputs controlling inputs rotation of the upper carriage, boom lift, boom tilt and bucket. The resulting action vector takes the form: a = [uuc ulb uda ub]T (19) where uuc is the upper carriage input, ulb the lift boom input, uda the dipper arm input and ub the bucket rotation input. The input uuc controls the hydraulic motor slew, which rotates the 214002 VOLUME 8, 2020\n[page 6] I. Kurinov et al.: Automated Excavator Based on RL and Multibody System Dynamics FIGURE 4. Topological structure of the excavator model with hydraulic actuators. FIGURE 5. Hydraulic circuit schematics of the excavator. upper carriage around Y-axis relative to the under carriage. The lift boom input ulb controls the boom lift hydraulic cylin- ders, which rotates the lift boom around Z-axis relative to the upper carriage. The dipper arm input uda controls the dipper arm hydraulic cylinder, which rotates the dipper arm around Z-axis relative to the lift boom. The bucket input ub controls the hydraulic cylinder, which rotates the bucket around Z-axis relative to the dipper arm, please see Figure 4. For evaluation of the performed actions, the agent must access the observation vector and reward function. The observation vector of the excavator model is different for each reward function type but contains the same parameters groups, which are set as bodies, task and goal position. The first group consists of the global coordinates of the bodies in meters and the rotations of the bodies in Euler parameters. In the case of an excavator, this group includes the upper carriage, lift boom, tilt boom and bucket bodies. Task-related observations are represented by values from sensors. Normal- ized mass in the hopper and bucket, a number of collisions of the bucket with the hopper are used as the goal group. Com- bination of these groups results to the following observation vector: s = \u0002 t mh mb lt−1 lt 1l rT rT g nc ng \u0003T (20) where, t is elapsed time, mh is normalized mass in the hopper, mb is normalized mass in the bucket, lt−1 is the distance to current goal on the previous time step, lt is the distance to the goal, 1l is the difference between lt and lt−1, r is position vector of bodies, rg is position vector of goals, nc is the number of the collisions during one episode and ng is the number of the goals already reached. The r vector has a size of 28, which is equal to the number of positions and rotations of the 4 main moving bodies. It consists of: r = \u0002 rT b1 rT b2 rT b3 rT b4 \u0003T (21) where, rbn is the vector of the body n positions."}], "source_title": "Automated Excavator Based on Reinforcement Learning and Multibody System Dynamics", "source_doi": "10.1109/access.2020.3040246", "source_url": "https://ieeexplore.ieee.org/ielx7/6287639/6514899/09268069.pdf", "source_license": "cc-by-sa", "zotero_item_key": "7F2IQNFD"}
{"id": "377dfc0187ae3078", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Automated Excavator Based on Reinforcement Learning and Multibody System Dynamics; DOI: 10.1109/access.2020.3040246; license: cc-by-sa."}, {"role": "assistant", "content": "The reinforcement learning method applies Proximal Policy Optimization with Covari- ance Matrix Adaptation (PPO-CMA) algorithm [15]. The agent is operating the excavator model that was based on the use of the multibody system dynamics. The current study employs the semi-recursive multibody approach, in which relative joint coordinates are used to minimize the number of differential equations [16]–[19]. The excavator is actuated by hydraulics. In this study, the hydraulic systems of the exca- vator are modeled using lumped fluid theory, in which the hydraulic system is discretized to volumes [20], [21]. To sim- ulate real-life tasks of the excavator, deformable ground model based on a combination of the cellular automata and particle method is implemented [22]–[25]. The agent and the multibody solver are connected by ZeroMQ [26] asyn- chronous messaging library, where reinforcement learning algorithm is the server side and simulation environment is the client side. The objective of this study is to computationally combine efficient multibody simulations with the reinforcement learn- ing algorithm. This combination, in turn, enables the creation and training of an autonomous agent. In this way autonomous agents can be designed effectively without endangering per- sonnel or machinery. II. METHODS This study combines five elements. The first element is reinforcement learning implemented using the PPO-CMA algorithm [15]. The second element is a multibody system dynamics solver that uses a semi-recursive method [16]–[19]. The third method is the hydraulics simulation using lumped fluid theory [20], [21]. The fourth method is a deformable ground implemented in the multibody simulation by combin- ing cellular automata and particle methods [22]–[25]. The last element focuses on connecting the RL algorithm with the multibody simulation via a custom application program interface (API) in Python based on ZeroMQ [26]. VOLUME 8, 2020 213999\n[page 3] I. Kurinov et al.: Automated Excavator Based on RL and Multibody System Dynamics A. REINFORCEMENT LEARNING AND PROXIMAL POLICY OPTIMIZATION WITH COVARIANCE MATRIX ADAPTATION ALGORITHM Reinforcement learning is a machine learning paradigm in which learning is realized by a trial and error process. Unlike the other two main machine learning techniques, supervised learning and unsupervised learning, RL does not require a predefined dataset and uses the environment directly for learning. The environment and agent are usually simulation models that enable observation – a vector of data about the environment and reward – which is the score of a performed action. The agent is the part of the simulation capable of taking action. The action is a set of inputs for interaction with the environment. The agent interacts with an environment by providing an action vector, getting an observation vector and receiving evaluation via a reward function. Reinforcement learning algorithms can be divided into model-based and model-free algorithms [27]. The differ- ence between model-based and model-free approaches is the knowledge of the agent about the reward function and transition matrix which maps observations to actions. In model-based approaches, the transition matrix and reward function define the model of the world. On the other hand, in model-free methods, the agent does not have knowledge about the transition matrix and reward function and, thus, the agent must estimate state-action pairs according to pre- vious experience. This research focuses on application of a model-free policy optimization method. The policy defines the behavior of the agent. In reinforcement learning, the policy is parametrized, e.g., by weights and biases of the neural network, which can be adjusted to change the behavior of the agent. A common policy optimization method is Proximal Policy Optimization (PPO), which has been shown to be able to operate in complex environments [28]. The basic idea of PPO is to optimize the policy with a gradient and limit changes to the area of sampled actions [28]. Nevertheless, PPO has a tendency to slow down or get stuck at local optima, because of prema- ture shrinkage of the exploration area [15]. The algorithm called Proximal Policy Optimization with Covariance Matrix Adaptation was designed to alleviate this problem [15]. The PPO-CMA algorithm is presented in the PPO-CMA study [1]. B."}], "source_title": "Automated Excavator Based on Reinforcement Learning and Multibody System Dynamics", "source_doi": "10.1109/access.2020.3040246", "source_url": "https://ieeexplore.ieee.org/ielx7/6287639/6514899/09268069.pdf", "source_license": "cc-by-sa", "zotero_item_key": "7F2IQNFD"}
{"id": "08460695eced116f", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: How to evaluate performance of prediction methods? Measures and their interpretation in variation effect analysis; DOI: 10.1186/1471-2164-13-s4-s2; license: cc-by."}, {"role": "assistant", "content": "Further analyses were made about variations located in different protein secondary structural elements, on the surface or in the core of a protein, and according to protein structure type. The conclusion was that even at best, the predictions were only moderately accurate (~60%) and significant improvements would be needed. The correlation of the methods was poor. In another study six programs includeing CC/PBSA, EGAD, FoldX, I-Mutant2.0, Rosetta, and Hunter were compared [3]. The dataset contained 2156 single varia- tions from ProTherm. The goal of the study was to com- pare the performance of the methods in ΔΔG prediction. Thus, they did not directly predict the effect on protein function, just the extent of free energy change. The only measure used was correlation between the experimental and predicted ΔΔG values. The ability of Dmutant, two versions of I-Mutant 2.0, MUpro, and PoPMuSiC to detect folding nuclei affected by variations has been evaluated [34]. The dataset con- tained 1409 variations from the ProTherm and some methods were tested with the same data which they had been trained. They used only correlation coefficients as quality measures. The best being in the range of ~0.5. The performance of structure-based stability preditors, Dmutant, FoldX, and I-Mutant 2.0, were investigated with data for two proteins. There were 279 rhodopsin and 54 bacteriorhodopsin variations [35]. The best prediction accuracy for the rhodopsin dataset was <0.60, while it was somewhat greater for the bacteriorhodopsin dataset. Splice site predictors mRNA maturation is a complex process, which may be affected by variations in many steps. Prediction beha- viour of nine systems, GenScan, GeneSplicer, Human Splicing Finder (HSF), MaxEntScan, NNSplice, Splice- Port, SplicePredictor, SpliceView and Sroogle was tested [4]. The test dataset contained altogether 623 variations. The first dataset contained 72 variations that affect the four invariant positions of 5’ and 3’ splice sites. The second one included 178 variations either localized at splice sites in non-canonical positions, distant intronic variations, and short distance variations. The third set of 288 exonic varia- tions included 10 exonic substitutions that activate a cryp- tic splice site. In the fourth dataset were negative controls, altogether 85 variations without effect on splicing. The results contain just the numbers of predicted cases and the percentage of correct ones, thus detailed analysis of the merits of the methods cannot be made. The authors recommended some programs but stated that the in silico predictions need to be validated in vitro. Checklist for method developers and users This checklist is provided to help when comparing and measuring performance of predictors and when selecting a suitable one. These are items that method developers should include in articles, or as supplement to articles, as they enable effective comparison and evaluation of the performance of predictors. Items to check when estimating method performance and comparing performance of different methods: - Is the method described in detail? - Have the developers used established databases and benchmarks for training and testing (if available)? - If not, are the datasets available? - Is the version of the method mentioned (if several versions exist)? - Is the contingency table available? - Have the developers reported all the six performance measures: sensitivity, specificity, positive predictive value, negative predictive value, accuracy and Matthews correlation coefficient. If not, can they be calculated from figures provided by developers? - Has cross validation or some other partitioning method been used in method testing? - Are the training and test sets disjoint? - Are the results in balance e.g. between sensitivity and specificity? - Has the ROC curve been drawn based on the entire test set? - Inspect the ROC curve and AUC. - How does the method compare to others in all the measures? - Does the method provide probabilities for predictions? Acknowledgements This work was supported by the Sigrid Jusélius Foundation, Biocenter Finland and the Competitive Research Funding of Tampere University Hospital."}], "source_title": "How to evaluate performance of prediction methods? Measures and their interpretation in variation effect analysis", "source_doi": "10.1186/1471-2164-13-s4-s2", "source_url": "https://bmcgenomics.biomedcentral.com/counter/pdf/10.1186/1471-2164-13-S4-S2", "source_license": "cc-by", "zotero_item_key": "LKJJ9WPZ"}
{"id": "13cf6c96959b8288", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Wheel Loader Scooping Controller Using Deep Reinforcement Learning; DOI: 10.1109/access.2021.3056625; license: cc-by-sa."}, {"role": "assistant", "content": "The sequence of state-action pairs defines the trajectory τ = (s0, a0, s1, a1 . . . . . . sH, aH) of length H. The return Rγ t is the total discounted reward for the initial state over the trajectory Rγ t = PH i=t γ i−tr (si, ai). The goal in reinforcement learning is to learn a policy which maximizes the expected return from the initial distribution J = E \u0002 Rγ 1 |π \u0003 . The state-value function is defined as the expectation value of the return over all allowed trajectories from a specific state s, V π (st) = Eπ \u0002 Rγ t |st \u0003 . While the action-value function describes the expected return after tak- ing an action at in state st, Qπ (st, at) = Eπ \u0002 Rγ t |st, at \u0003 . DDPG is an actor-critic off-policy gradient algorithm that implements a stochastic behavior policy and estimates a deterministic target policy [16]. A schematic of the pro- cedure can be seen in Fig. 2. DDPG primarily uses two neural networks in its learning phase, one for the actor (policy network) and one for the critic (value network), with weights θµ and θQ, respectively. The actor network is used to approximate the optimal policy whereas the critic utilizes the value-based approach to estimate the value of state-action pairs. The actor function µ (st|θµ) deterministically maps states for specific actions. The critic’s output is the estimated Q-value of the current state and action provided by the actor Q \u0000st, at|θQ\u0001 . These networks compute action predictions for the current state and generate the Temporal-Difference (TD) error at each time step while using a set of target networks µ′ (st) , Q′ (st, at), with weights θµ′ and θQ′ respectively. The critic’s loss function L, and target yt are computed from the sum of the immediate reward and the outputs of the target actor and critic networks: yi = ri + γ Q′ h si+1, µ′ \u0010 si+1|θµ′\u0011 |θQ′i (1) L = 1 N X i \u0010 yi −Q \u0010 si, ai|θQ\u0011\u00112 (2) VOLUME 9, 2021 24147\n[page 4] O. Azulay, A. Shapiro: Wheel Loader Scooping Controller Using Deep Reinforcement Learning FIGURE 3. Architecture of the Komodo robot with our arm mechanism. The critic is updated by minimizing the loss and the actor is updated by applying the chain rule using the sampled policy gradient: ∇θµJ ≈1 N X i × h ∇aQ \u0010 s, a|θQ\u0011 |s=si,a=µ(si|θµ)∇θµµ \u0000s|θµ\u0001 |s=si i (3) IV. SYSTEM DESIGN To investigate bucket-scooping processes, we developed an arm mechanism that we attached to a Komodo robot (see Fig. 3). The robot is a 40kg weight fully Robot Operating System (ROS) supported skid-steer robot that includes an Intel NUC i7 CPU, SICK LMS1 range laser scanner, an Asus Xtion pro depth camera, and 6-axis IMU. To control the robot motion, we used the Komodo built-in Roboteq differential driver for the motor controllers. For the loading mechanism, earth-moving mechanisms typically consist of a robotic arm controlled by hydraulic pistons and a bucket. In this work, we developed an arm mechanism that resembles to an indus- trial skid-steer loader that consists of three main parts: 1) a base link, 2) an arm assembly, and 3) a bucket, as shown in Fig. 3. The mechanism is constructed primarily from strong lightweight 6061 aluminum parts that can withstand a bucket load of up to 12[Kg]. The mechanism is controlled by two parallel systems of linear actuators with a built-in position feedback potentiometer. The arm’s linear actuators are responsible for the lift motion, whereas the bucket’s linear actuators generate its tilting motion. To control tilt and lift motion, we used a Proportional Integral Derivative controller (PID), employing feedback from the linear actuator’s poten- tiometer. The system is operated using the Robot Operation System (ROS) interface. The electrical system is depicted in Fig. 4, and the com- ponents are listed in Table 1. We built the electric cir- cuit to be powered by the robot’s power source, whereas the components with other power requirements utilized FIGURE 4. The electrical scheme of the arm mechanism. TABLE 1. Arm mechanism components. voltage converters. We divided the control system into two levels. The higher level was composed of the PID application on the tilt and lift motion, communication initializing, and syncing between the various channels using ROS. The lower- level controller was composed of two slave nodes that sub- scribe to the data published by the main controller channels."}], "source_title": "Wheel Loader Scooping Controller Using Deep Reinforcement Learning", "source_doi": "10.1109/access.2021.3056625", "source_url": "https://ieeexplore.ieee.org/ielx7/6287639/9312710/09344588.pdf", "source_license": "cc-by-sa", "zotero_item_key": "IDZBIHT7"}
{"id": "e640fef4bc77ec57", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Detection of Ionospheric Scintillation Based on XGBoost Model Improved by SMOTE-ENN Technique; DOI: 10.3390/rs13132577; license: cc-by-sa."}, {"role": "assistant", "content": "The wideband antenna was split into several ports that were connected to a commercial ISM receiver, Septentrio PolaRx, and SDR-based RF front ends, respectively. Scintillation-related mea- surements including channel correlation values were continuously collected by the ISM receiver, and relevant scintillation indices and event indicators were calculated simulta- neously [10]. Furthermore, the indicators were to be compared with the threshold values preset before data collection to trigger the data server to record raw IF samples generated by the SDR front ends [38,39]. It should be mentioned that only data affected by the natural ionospheric scintillation phenomenon were to be recorded, to save a large amount of memory. These data could be utilized in research on optimal algorithms with respect to the ISM receiver, as well as providing a database for the analysis of strong scintillation characteristics. It is worth mentioning that, in the process of the acquisition and tracking, the DLL and PLL pull-in noise bandwidths were set as 2 Hz and 25 Hz, respectively, while the pull-in time was set as 500 ms. Once the signals were locked and the tracking loop was kept stable, the DLL and PLL pull-in noise bandwidths were reset to 1 Hz and 10 Hz, respectively, to reduce the influence of noise and other sources of interference. Natural Ionospheric Scintillation Events Wideband Antenna Splitter Commercial Scintillation Receiver(Septentrio PolaRx) RF Front End 1 RF Front End 2 RF Front End N · · · SDR Front-ends Circular Buffer 1 Circular Buffer 2 Circular Buffer N · · · Event Data Buffers Autonomous Scintillation Monitoring and Event Trigger Data Collection and Control Server Data Storage Figure 2. Architecture of the scintillation event-driven data collection system developed in the equatorial region. The commercial ISM Septentrio PolaRx is used to collect relevant navigation data and trigger SDR-based RF front ends when natural scintillation events occur.\n[page 6] Remote Sens. 2021, 13, 2577 6 of 22 A dataset of GPS L1 C/A signals was selected for the study. It was recorded from the GTEC free front ends, which were configured to collect zero-frequency data with 8-bit- resolution samples at a 20 MHz complex sampling rate. The collection period lasted from March 2013 to February 2015, covering the latest peak period for solar activity. Furthermore, these available data were processed by the SDR receiver advanced by a combination coherent/con-coherent integration acquisition algorithm to enhance the acquisition and tracking performance under the strong ionospheric scintillation environment. Overall, 45 segments of one-hour data were used in the following research, with two or three satellites’ data selected in each segment. The I/Q values were sampled at 1000 Hz, while the sample rate was set to 1 Hz according to the size of the shifting window. Based on three levels of scintillation intensity, Table 1 presents the information of the training dataset and testing dataset used in the following research. Table 1. The distribution information of training dataset and testing dataset on three classes. Dataset Weak (Class 0) Moderate (Class 1) Strong (Class 2) Total Ratio Training 155,892 136,343 38,515 349,559 4.05:3.54:1 Testing 18,680 14,496 5665 38,841 3.30:2.56:1 2.3. Feature Extraction The features play a decisive role in the classification model. There are different sets of features producing different results for the detection accuracies, while the set of signal- based features as well as corresponding combinations are proved to be optimal with the location-independent technique [16]. The raw GNSS signal measurements in phase (I) and quadrature-phase (Q) correlators were utilized as features, which were extracted from the software GNSS receiver at the end of the tracking process. In recent research [16], the feature set of \b⟨I⟩, ⟨Q⟩, I2 , Q2 , ⟨SI⟩, SI2 —namely, the average values of I and Q correlator outputs, I2, Q2, signal intensity SI and SI2—has been proved to show excellent performance for ionospheric scintillation detection by machine learning algorithms."}], "source_title": "Detection of Ionospheric Scintillation Based on XGBoost Model Improved by SMOTE-ENN Technique", "source_doi": "10.3390/rs13132577", "source_url": "https://www.mdpi.com/2072-4292/13/13/2577/pdf?version=1625212202", "source_license": "cc-by-sa", "zotero_item_key": "4YEIWMK3"}
{"id": "65e8ed30937ed285", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Automated Excavator Based on Reinforcement Learning and Multibody System Dynamics; DOI: 10.1109/access.2020.3040246; license: cc-by-sa."}, {"role": "assistant", "content": "[page 1] Received October 16, 2020, accepted November 15, 2020, date of publication November 24, 2020, date of current version December 10, 2020. Digital Object Identifier 10.1109/ACCESS.2020.3040246 Automated Excavator Based on Reinforcement Learning and Multibody System Dynamics ILYA KURINOV 1, GRZEGORZ ORZECHOWSKI 1,2, PERTTU HÄMÄLÄINEN 3, AND AKI MIKKOLA 1 1Department of Mechanical Engineering, LUT University, 53850 Lappeenranta, Finland 2Mevea Ltd., 53850 Lappeenranta, Finland 3Department of Computer Science, Aalto University, 02150 Espoo, Finland Corresponding author: Ilya Kurinov (ilya.kurinov@lut.fi) This work was supported in part by the European Union’s Horizon 2020 Research and Innovation Programme through the Marie Sklodowska-Curie Project under Grant 845600 (RealFlex) in part with Academy of Finland #316106. ABSTRACT Fully autonomous earth-moving heavy equipment able to operate without human intervention can be seen as the primary goal of automated earth construction. To achieve this objective requires that the machines have the ability to adapt autonomously to complex and changing environments. Recent developments in automation have focused on the application of different machine learning approaches, of which the use of reinforcement learning algorithms is considered the most promising. The key advantage of reinforcement learning is the ability of the system to learn, adapt and work independently in a dynamic environment. This article investigates an application of reinforcement learning algorithm for heavy mining machinery automation. To this end, the training associated with reinforcement learning is done using the multibody approach. The procedure used combines a multibody approach and proximal policy optimization with a covariance matrix adaptation learning algorithm to simulate an autonomous excavator. The multibody model includes a representation of the hydraulic system, multiple sensors observing the state of the excavator and deformable ground. The task of loading a hopper with soil taken from a chosen point on the ground is simulated. The excavator is trained to load the hopper effectively within a given time while avoiding collisions with the ground and the hopper. The proposed system demonstrates the desired behavior after short training times. INDEX TERMS Autonomous agents, discrete event dynamic automation systems, learning and adaptive systems, real-time simulation, multibody system dynamics, reinforcement learning, PPO-CMA. I. INTRODUCTION Mobile earth-moving machinery is widely used in the con- struction, forestry and mining industries. The operation of such machinery is challenging and often hazardous, because the machines are used in dangerous environments like open-pit mines and large infrastructure projects. From the performance perspective, machine operators are often the weakest link, as they introduce delays in the form of statutory rest periods, transportation to and from the worksite, and fatigue-induced errors. Consequently, many manufacturers of heavy machinery have begun to explore greater automation of machine operations. Tele-remote operation [1] is nowa- days possible, which eliminates the aspect of the dangerous environment, but other issues remain. To fully address the The associate editor coordinating the review of this manuscript and approving it for publication was Wei Zhang. challenges associated with heavy earth-moving operations, there is a need for fully automated and operator-independent machines. Use of autonomous earth-moving machinery can increase safety and productivity, but production and development of such equipment is a formidable task. The main chal- lenge facing earth-moving machinery automation is the com- plexity of the working environment. In mining and other earth-moving operations, the environment is characterized by diversity, variable geometry, the effects of the machine or other machines on the environment itself, the properties of the soil or media to be excavated, and external factors such as ambient temperature and climatic conditions. Control of machine operations using conventional methods is therefore a challenging enterprise [2]. The autonomous machine should be able to adapt to the sum of external factors and make decisions based on the task and state of the environment. 213998 This work is licensed under a Creative Commons Attribution 4.0 License."}], "source_title": "Automated Excavator Based on Reinforcement Learning and Multibody System Dynamics", "source_doi": "10.1109/access.2020.3040246", "source_url": "https://ieeexplore.ieee.org/ielx7/6287639/6514899/09268069.pdf", "source_license": "cc-by-sa", "zotero_item_key": "7F2IQNFD"}
{"id": "5fbfb708425ba73a", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Detection of Ionospheric Scintillation Based on XGBoost Model Improved by SMOTE-ENN Technique; DOI: 10.3390/rs13132577; license: cc-by-sa."}, {"role": "assistant", "content": "Feature matrix Training data (scintillation data) Feature extraction Manual visual inspection Feature matrix Label vector SMOTE-ENN resampling XGBoost algorithm Trained model Cross validation Testing data (scintillation data) Feature extraction Predictable model Predicted labels Manual visual inspection Manual labels Performance evaluation New dataset Resampling Training Predicting Raw dataset Figure 5. The process flow of XGBoost detection model improved by SMOTE-ENN resampling technique. 4. Results A series of experiments were conducted to evaluate the performance of the proposed method and compare it with that of the decision tree and random forest algorithms, which have shown high validation accuracy in previous research [16]. The evaluation criterion with which the results were quantitatively analyzed is introduced. Based on the validation\n[page 11] Remote Sens. 2021, 13, 2577 11 of 22 phase, the overall accuracy of three different machine learning methods in the condition of different tree depths was compared, determining the complexity, computational efficiency and performance of these models to a large extent. Next, different resampling methods were implemented on the training dataset, comparing their performance in improving the detection of the minority class. Finally, experiments on training sets with different ratios of the three classes were carried out to evaluate the influence of imbalanced data on scintillation detection, while previous similar experiments were conducted on testing sets. 4.1. Evaluation Criterion As one of the statistical tools most commonly used to evaluate the performance of detection algorithms, the confusion matrix, aimed at detection for three classes, was introduced. According to that, the corresponding metrics of overall accuracy, precision, recall and F1 score could be calculated to assess the performance of the detection models. Table 3 shows the example of the confusion matrix on three classes. Table 3. The confusion matrix on three classes. Class Prediction 0 1 2 Truth 0 N00 N01 N02 1 N10 N11 N12 2 N20 N21 N22 In Table 3, Nij (i, j = 0, 1, 2) refers to the number of samples corresponding to the truth class i and prediction class j. According to that, the relative evaluation indicators on three classes can be described as follows: accuracy = N00 + N11 + N22 N00 + N10 + N20 + N01 + N11 + N21 + N02 + N12 + N22 (11)      precision0 = N00 N00+N10+N20 precision1 = N01 N01+N11+N21 precision2 = N02 N02+N12+N22 (12)      recall0 = N00 N00+N01+N02 recall1 = N10 N10+N11+N12 recall2 = N02 N20+N21+N22 (13)        F1 −score0 = 2 ∗precision0 ∗recall0 precision0+recall0 F1 −score1 = 2 ∗precision1 ∗recall1 precision1+recall1 F1 −score1 = 2 ∗precision1 ∗recall1 precision1+recall1 (14) where the precision value represents the correct predicted positive ratio, and the recall value refers to the percentage of correctly predicted positive events among real positive events. Based on that, the weighted average of the precision and recall F1 score is defined as the F1 score. 4.2. Accuracy Evaluation on Cross Validation As an important parameter for algorithms based on the decision tree, random forest and XGBoost, the tree depth determines the complexity and performance of models. Excessive depth might increase the validation accuracy, but can also lead to increased model complexity and greater computational loads, causing overfitting. Therefore, an appropriate tree depth is vital for the training model. Based on cross-validation, Figure 6 shows the mean accuracies across different tree depths and detection methods. Of the overall 349,559 points of the training set, 100,000 entries were randomly selected and used\n[page 12] Remote Sens. 2021, 13, 2577 12 of 22 in the validation experiment to reduce the running time. The accuracies of XGBoost are marked, ranging from 1 to 30, and compared with those of the decision tree and random forest, which presented high validation accuracies in previous research [16]. 0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 Tree depth 76 80 84 88 92 96 100 XGBoost Random Forest Decision Tree Figure 6. The mean accuracy of 10-fold cross validation based on different tree depths and detection methods. In the overall training dataset mentioned in Table 1, 100,000 samples were selected randomly and used in the validation process, and the remaining data were used as training samples."}], "source_title": "Detection of Ionospheric Scintillation Based on XGBoost Model Improved by SMOTE-ENN Technique", "source_doi": "10.3390/rs13132577", "source_url": "https://www.mdpi.com/2072-4292/13/13/2577/pdf?version=1625212202", "source_license": "cc-by-sa", "zotero_item_key": "4YEIWMK3"}
{"id": "c00c987ba2a3d3a8", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Spectral Angle Mapping and AI Methods Applied in Automatic Identification of Placer Deposit Magnetite Using Multispectral Camera Mounted on UAV; DOI: 10.3390/min12020268; license: cc-by-sa."}, {"role": "assistant", "content": "Rauhala, A.; Tuomela, A.; Davids, C.; Rossi, P. UAV Remote Sensing Surveillance of a Mine Tailings Impoundment in Sub-Arctic Conditions. Remote Sens. 2017, 9, 1318. [CrossRef] 14. Martelet, G.; Gloaguen, E.; Døssing, A.; Lima Simoes da Silva, E.; Linde, J.; Rasmussen, T.M. Airborne/UAV Multisensor Surveys Enhance the Geological Mapping and 3D Model of a Pseudo-Skarn Deposit in Ploumanac’h, French Brittany. Minerals 2021, 11, 1259. [CrossRef] 15. Saha, D.; Annamalai, M. Machine Learning Techniques for Analysis of Hyperspectral Images to Determine Quality of Food Products: A Review. Curr. Res. Food Sci. 2021, 4, 28–44. [CrossRef] [PubMed] 16. Girouard, G.; Bannari, A.; Harti, A.E.; Desrochers, A. Validated Spectral Angle Mapper Algorithm for Geological Mapping: Comparative Study between Quickbird and Landsat-TM. 6. 2014. Available online: https://www.researchgate.net/publication/ 228799788_Validated_spectral_angle_mapper_algorithm_for_geological_mapping_comparative_study_between_QuickBird_ and_Landsat-TM (accessed on 4 July 2021). 17. Shafri, H.Z.M.; Suhaili, A.; Mansor, S. The Performance of Maximum Likelihood, Spectral Angle Mapper, Neural Network and Decision Tree Classifiers in Hyperspectral Image Analysis. J. Comput. Sci. 2007, 3, 419–423. [CrossRef] 18. den Hartog, D.; Harlaar, J.; Smit, G. The Stumblemeter: Design and Validation of a System That Detects and Classifies Stumbles during Gait. Sensors 2021, 21, 6636. [CrossRef] [PubMed] 19. Chauhan, D.; Anyanwu, E.; Goes, J.; Besser, S.A.; Anand, S.; Madduri, R.; Getty, N.; Kelle, S.; Kawaji, K.; Mor-Avi, V.; et al. Comparison of Machine Learning and Deep Learning for View Identification from Cardiac Magnetic Resonance Images. Clin. Imaging 2022, 82, 121–126. [CrossRef] [PubMed] 20. Kobayashi, S.; Ikuta, K.; Sugimoto, R.; Honda, H.; Yamada, M.; Tominaga, O.; Shoji, J.; Taniguchi, M. Estimation of submarine groundwater discharge and its impact on the nutrient environment at Kamaiso beach, Yamagata, Japan. Nippon. Suisan Gakkaishi 2019, 85, 30–39. [CrossRef] 21. Nguyen, H.H.; Carter, A.; Hoang, L.V.; Vu, S.T. Provenance, Routing and Weathering History of Heavy Minerals from Coastal Placer Deposits of Southern Vietnam. Sediment. Geol. 2018, 373, 228–238. [CrossRef] 22. Beretta, F.; Rodrigues, A.L.; Peroni, R.L.; Costa, J.F.C.L. Automated Lithological Classification Using UAV and Machine Learning on an Open Cast Mine. Appl. Earth Sci. 2019, 128, 79–88. [CrossRef] 23. Ono, M. Parallax Error Correction Techniques by Image Matching for ASTER/SWIR Band-to-Band Registration. In Proceedings of the Platforms and System, Rome, Italy, 9 January 1995; pp. 18–27. [CrossRef] 24. Laurence, S.J. On Tracking the Motion of Rigid Bodies through Edge Detection and Least-Squares Fitting. Exp. Fluids 2012, 52, 387–401. [CrossRef] 25. Singh Kushwah, J.; Kumar, A.; Patel, S.; Soni, R.; Gawande, A.; Gupta, S. Comparative Study of Regressor and Classifier with Decision Tree Using Modern Tools. Mater. Today Proc. 2021, S2214785321076574. [CrossRef] 26. Mohammadi, M.; Rezaei, J. Ensemble Ranking: Aggregation of Rankings Produced by Different Multi-Criteria Decision-Making Methods. Omega 2020, 96, 102254. [CrossRef] 27. Haixiang, G.; Yijing, L.; Shang, J.; Mingyun, G.; Yuanyue, H.; Bing, G. Learning from Class-Imbalanced Data: Review of Methods and Applications. Expert Syst. Appl. 2017, 73, 220–239. [CrossRef] 28. Gholami, R.; Moradzadeh, A.; Yousefi, M. Assessing the Performance of Independent Component Analysis in Remote Sensing Data Processing. J. Indian Soc. Remote Sens. 2012, 40, 577–588. [CrossRef]"}], "source_title": "Spectral Angle Mapping and AI Methods Applied in Automatic Identification of Placer Deposit Magnetite Using Multispectral Camera Mounted on UAV", "source_doi": "10.3390/min12020268", "source_url": "https://www.mdpi.com/2075-163X/12/2/268/pdf?version=1645347690", "source_license": "cc-by-sa", "zotero_item_key": "VPIE7856"}
{"id": "fbb27b457d095cb8", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Reinforcement learning based process optimization and strategy development in conventional tunneling; DOI: 10.1016/j.autcon.2021.103701; license: cc-by."}, {"role": "assistant", "content": "11 it can be seen that within individual sequences of top heading or bench excavation, the agent alternates between supported and unsupported blasts which shows that it has successfully learned to avoid excessive use of support ahead of the face. 6. Conclusion and implications for digital tunneling Before drawing conclusions from the experiments, it should be pointed out that in this initial study, the agent’s possibilities to find creative and not yet imagined solutions to real world tunneling problems are confined to the given set of actions and the taken simplifications (see Section 2.1). We therefore see it as a success that the agent optimizes the given scenario and finds strategies that are comparable to current practices in tunneling. As given in the introduction, the current study should serve as a base for future developments of RL in tunneling. Consequently, optimizations that improve the current practices of “real life” tunneling are to be expected from future studies. In the experiments of the previous chapter, we can observe that the agents have found policies that minimize the overall amount of neces­ sary blasts, minimize changes between top heading and bench excava­ tion, favor long over short advance lengths and minimize the use of face support. These strategies show that the agents have learned to work in an efficient and economically optimized way. We see similarities in this RL-based tunneling strategy to real world tunneling paradigms like the NATM [49] which uses partial excavation to minimize the necessary support. Minimizing the number of changes between top heading or bench excavation while at the same time not exceeding a too long dis­ tance between them is also part of NATM tunneling as this optimizes construction site logistics on the one hand and safety requirements on the other. The found strategies that rely on support ahead of the face show similarities to the “Adeco” method [25] which uses heavy support installation and long advance lengths to deal with the encountered rockmass conditions. While adhering to safety requirements is impera­ tive, most “real life” tunneling methods work towards minimizing the number of necessary blasts which is a policy that was found by all the agents. TunnRL (see Section 2) has shown that it is not only a functioning environment/simulation of conventional tunneling, but also that a RL agent can successfully interact with it and learn optimized and inno­ vative strategies that seem realistic compared to real world tunnel excavation. At the same time, we see the challenges of computational instabilities and trial and error approach in the process of developing well-functioning models, highlighting this early stage in RL for tunneling. Clearly both the reward system, the rockmass-environment and the action-system has room for improvement. Still we see a signif­ icant potential in the TunnRL-concept: firstly, for an on-face decision support system in a further developed and more realistic version, and secondly as a first step to more advanced automation in underground Table 5 Statistics of the training runs presented in Fig. 7. Values were computed from the 500-episode sliding window average and the episode of the respective value is given in parenthesis behind it. Row-wise best performances are highlighted in grey. G.H. Erharter et al.\n[page 10] Automation in Construction 127 (2021) 103701 10 construction. Where many of the developments of tunnel processes today address the automation of small-scale processes in the tunnel or at the excavation face, a further developed TunnRL could be part of the main controlling mechanism that operates the overall tunnel construc­ tion site. Albeit the agents in this simulation were not able to find new and undiscovered strategies of tunnel excavation (see previous chapter), we see the fact that completely untrained agents are able to find tunneling strategies which are comparable to “real world tunneling” as a proof of concept that RL is successfully applicable to this kind of prob­ lem. The main goal of the study is therefore fulfilled, and future studies will work towards giving the agent more capabilities and increasing the environment’s realism. This will ultimately pave the way for an opti­ mized decision finding process in sequential tunneling. 7. Outlook Albeit we designed TunnRL in a practice related context, there are numerous improvements to make the framework more realistic and more robust."}], "source_title": "Reinforcement learning based process optimization and strategy development in conventional tunneling", "source_doi": "10.1016/j.autcon.2021.103701", "source_url": "https://www.sciencedirect.com/science/article/pii/S0926580521001527", "source_license": "cc-by", "zotero_item_key": "RXHUVV3X"}
{"id": "ff7444d4894159bd", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Reinforcement learning based process optimization and strategy development in conventional tunneling; DOI: 10.1016/j.autcon.2021.103701; license: cc-by."}, {"role": "assistant", "content": "A copy of the agent is saved after every 1000 episodes. To observe how big the differences are between individual training runs, we trained several identical DQN agents in the above described environment. After every episode, 21 parameters are saved to monitor the training progress. Below, the parameters that are mentioned in the paper are given (see the code “A_utilities.py” in Appendix 1 for a list of all recorded parameters): ▪ number of the current episode ▪ cumulative reward of the episode ▪ current value of ε ▪ the number of face instabilities of the whole episode ▪ average loss of the DQN agent throughout the episode ▪ number of moves/blasts that were required to finish the episode/to reach a terminal state ▪ 8 counters for how many times each of the actions of Table 3 were used; the goal of these counters is to see if the agent favors some actions over others and to detect “strategy changes” 5. Experiments In this section we present five training paths of exemplary agents and discuss the different strategies they found to deal with the given task. In Fig. 7, recordings of the agents’ training are given which shows that each agent has found a unique solution, and all training paths are substan­ tially different from one another. Comparing the five agents to one another, all of them were able to increase the cumulative reward per episode above 200 within 10,000 episodes (Fig. 7 first row). After around 30,000 episodes, first differ­ ences arise where first the cumulative reward of agent 2 and then of agent 3 started to stagnate. The reward of agent 1 stagnates at around 80,000 episodes. The cumulative rewards of the agents 4 and 5 kept on increasing, whereas agent 5 became instable after around 75,000 epi­ sodes and agent 4 reached the maximum reward at the end of the 120,000 training episodes. Right from the start, all agents started to use long advance lengths as a means to decrease the number of blasts/moves per episode as this is an effective way to maximize the achievable reward (Fig. 7 second row). Where the agents 1, 4 and 5 all took a similar strategy that aims at continuously minimizing the blasts/moves per episode throughout the whole training, agents 2 and 3 reached a minimum of around 120 blasts/ moves per episode after around 30,000 episodes. This correlates well with the stagnating rewards after 30,000 episodes of agents 2 and 3 as described above. The biggest differences in training paths can be observed with Table 4 Rewards that the agent receives from the environment in response to its actions. Reward (points) Description Condition tl * 3 reward for achieving breakthrough posth ≥tl and posbi ≥tl tl * 3 * -1 penalty for a timeout if number of moves in current episode is >200 −6 penalty for using the wrong excavation sequence, i.e. the bench is driven further ahead than the top heading posbi > posth −5 penalty for unstable tunnel face conditions determined from face pressure pf as evaluated by Eq. (3) (see below) pf ≥0 −4 penalty for changing from top heading to bench excavation or vice versa, as this usually involves a delay of the excavation due to logistics if current a ≥200 and prev. a < 200 or if current a < 200 and prev. a ≥200 −3 penalty for a too far distance between top heading and bench (see Section 3.1) if posth – posbi > distmax (initially set to 50 m) −2 penalty for using face support as this consumes additional time and resources if a = 112; a = 152; a = 202; a = 222 −1 penalty for every other move that does not meet any of the above conditions no other condition is fulfilled G.H. Erharter et al.\n[page 8] Automation in Construction 127 (2021) 103701 8 respect to face instabilities, where only the agents 2 and 3 started to actively decrease the number of face instabilities (Fig. 7 third row) early in the training process (within 10,000 episodes). In contrast to that, the number of face instabilities increased within the first 10,000 episodes for the other agents and then only decreased slowly throughout the rest of the training. The best performing agent 4 shows a remarkable trend of stagnating face instabilities until around 80,000 episodes, followed by a decrease towards the end of training. While the goal was to let all agents train for 120,000 episodes, the training process of the agents 2 and 5 became unstable after around 75,000 and 85,000 episodes respectively, which led to the abortion of training after it could be observed that the agent would not recover from this."}], "source_title": "Reinforcement learning based process optimization and strategy development in conventional tunneling", "source_doi": "10.1016/j.autcon.2021.103701", "source_url": "https://www.sciencedirect.com/science/article/pii/S0926580521001527", "source_license": "cc-by", "zotero_item_key": "N5Q33JID"}
{"id": "bc06dd9a6294fe17", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Wheel Loader Scooping Controller Using Deep Reinforcement Learning; DOI: 10.1109/access.2021.3056625; license: cc-by-sa."}, {"role": "assistant", "content": "Another drawback is that the majority of work done to automate scooping processes is based on data recorded by expert operators, which limits the efficiency to the level of the operator’s skills [1]. Since the scooping cycle is a repetitive and dynamic task involv- ing intricate interactions with the environment, here a new strategy is put forward to solve the problem of autonomous bucket scooping automation for wheel loaders using Rein- forcement Learning (RL) methods. RL is a machine learning technique that has the potential to enable robots to learn large repertoires of behavioral skills with minimal human intervention through trial and error. The main advantage of RL is that it does not require a predefined control structure and can explore the environment to find a good policy to follow, thus mitigating the shortcoming of having to explicitly derive the interconnection with the environment. Deep Rein- forcement Learning (DRL) methods have been successfully applied in many complex robotic tasks from manipulation [11] to autonomous vehicles [12], and aerial applications [13]. However, practical real-world applications of RL are rel- atively rare since they often require unrealistic learning times, high sample complexity, and can potentially damage the robot [14], [15]. This article focuses on developing a scooping motion plan- ner that learns transferable scooping policies solely under simulation. We show that an earthmoving machine can be fully controlled using DRL methods trained solely in simula- tion and then deployed on a real Unmanned Ground Vehi- cle (UGV). The complete system architecture is depicted in Fig. 1. To further increase the robustness of the sys- tem and narrow the reality gap, we defined a compact design of the observation space and performed detailed system identification. The controller is trained using the model-free, actor-critic Deep Deterministic Policy Gradi- ent (DDPG) algorithm [16], which successfully completes scooping cycles without any user-provided demonstrations. We developed a ROS-Gazebo based open-source training environment (available at [17]) for the agent to learn scooping motions. To verify the learning algorithm and simulation viability, we deployed the controller in a real-world scenario. We also compared the performance of our DRL based con- troller against manual human control for an earthmoving cycle. The primary contribution of this paper is its novel end- to-end Neural Network (NN) based controller for a robotic wheel loader that autonomously scoops earth from a pile, with three degrees of freedom (DoF). The controller is trained solely via simulation without any prior knowledge using DRL. To the best of our knowledge, this is the first 3-DoF DRL scooping controller to learn from simulation without previous knowledge and be deployed on a real-life excavation machine. In addition, the development of an open source [17] training environment provides the research community with a tool to collect data and observe progress performance for excavation purposes. The remainder of this paper is structured as follows. Section II further discusses related work in this area, fol- lowed by a background review in Section III. Section IV describes the system design and Section V introduces the proposed scooping controller implementation. We demon- strate our simulation and report the experimental results in Section VI, followed by conclusions and suggested future research in Section VII. II. RELATED WORKS Most previous research on automating bucket-filling pro- cesses are based on data recorded by expert operators and implement a control system that can be generally divided into three main categories. The first is made up of position control algorithms for bucket motion trajectories, which are intended to maximize the volume scooped by the bucket [4], [5], [18], [19]. These approaches have been successful, but they rely on expert trajectories and do not generalize to different machine-pile environments. The second category is composed of compliance control algorithms, such as soil estimation based methods [6], [7] and force/torque based methods [20], [21]. Soil estimation methods predict the soil- tool interaction force and apply a heuristic-based motion by modifying the soil parameters in the program."}], "source_title": "Wheel Loader Scooping Controller Using Deep Reinforcement Learning", "source_doi": "10.1109/access.2021.3056625", "source_url": "https://ieeexplore.ieee.org/ielx7/6287639/9312710/09344588.pdf", "source_license": "cc-by-sa", "zotero_item_key": "IDZBIHT7"}
{"id": "45782556bca03a49", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Wheel Loader Scooping Controller Using Deep Reinforcement Learning; DOI: 10.1109/access.2021.3056625; license: cc-by-sa."}, {"role": "assistant", "content": "The drawbacks of soil estimation methods include the fact that they lack precision, fail to capture the rapidly changing properties in the environments or are non-real-time capable and thus not feasible/implementable [20]. Others combine these methods with a higher-level planner. For instance, a force-control tra- jectory controller with prioritized tasks [20], a coarse and fine planner that ensures equal performance over a large number of digs [22], or adding a disturbance observer to compensate for the difference between prediction and actual performance [21], [23]. These compliance control methods do not follow a desired trajectory but rather apply specific forces to the pile during the scooping motion. The third category is made up of control algorithms that employ a behavior-based approach for motion control, such as a rule-based algorithm that depends on the current phase and acts dynamically [8], [9]. Thus, most previous solutions to automate the scooping task (1) do not generalize to different machines or pile environ- ments (2) rely on prior knowledge of an expert operator and (3) require accurate models of the machine and therefore are susceptible to failure in the presence of modeling errors, wear and tear, and changing conditions [10]. This underscores the need for a generic automatic scooping solution that can be adapted to different scenarios. Few deep Learning methods have been implemented in earthmoving operations, especially in bucket-filling pro- cesses, mainly due to the task complexity. The chal- lenges include various problems such as simulating the environment, data collection, and soil-tool dynamics. Dadhich et al. [10], [24] demonstrated the implementation of a NN approach to control bucket motion during real-life 24146 VOLUME 9, 2021\n[page 3] O. Azulay, A. Shapiro: Wheel Loader Scooping Controller Using Deep Reinforcement Learning FIGURE 2. DDPG algorithm schematic. loading processes. They used data collected from an expert operator to train a time-delayed NN to perform a bucket- filling motion with two degrees of freedom (static wheels). Halbach et al. [25] implemented a shallow three-DoF NN controller for automated pile loading with a robotic wheel loader that used learning from the demonstration of sensor recordings. Both of these approaches were successful in terms of the bucket filling cycle, but nevertheless rely on the prior knowledge of expert operator data, which limits the efficiency to the operator’s skills which are not always available. III. PRELIMINARIES A. EARTHMOVING CYCLE A wide range of earth-moving machines are currently used in industry, the most common of which are wheel loaders and excavators. Excavators are used to dig into the ground whereas wheel loaders are used to load and transport exca- vated material. An earthmoving cycle is comprised of several steps: 1) navigating towards the pile; 2) scooping up the earth; 3) navigating to the dump location; and 4) dumping. This paper focuses on automating the bucket-filling process, which is the most complex phase. The filling of a bucket is a complex granular flow problem that can be divided into three general phases: approach, fill, and exit [3]. Each phase operates under different conditions in terms of vehicle feedback, terrain properties, and duration. In the first phase, approach, the front-loader moves towards the pile of earth, and the operator chooses the height and penetration angle of the front loader’s bucket. In the second phase, fill, there is a simultaneous change in the lift, tilt, and throttle actions to navigate the bucket tip through the earth pile while avoiding wheel slip and piston stall. The last phase, exit, involves tilting the bucket until the breakout, while the machine moves in reverse from the pile. B. REINFORCEMENT LEARNING In RL, the problem is typically characterized as a Markov Decision Processes (MDP), which defined by the tuple ⟨S, A, r, P, γ ⟩. S represents the state-space, and A is the action-space. Transitions between states are performed with transition probability p (st|st, at) ∈P, reward r (st, at) and a discount factor γ ∈[0, 1]. At each time step, the agent observes the current state st ∈S and takes action at ∈A according to the policy πθ, which can be stochastic or deter- ministic. While interacting with the environment, the system transitions to a new state st+1 ∈S and the agent receives a reward rt (st, at, st+1)."}], "source_title": "Wheel Loader Scooping Controller Using Deep Reinforcement Learning", "source_doi": "10.1109/access.2021.3056625", "source_url": "https://ieeexplore.ieee.org/ielx7/6287639/9312710/09344588.pdf", "source_license": "cc-by-sa", "zotero_item_key": "IDZBIHT7"}
{"id": "c777d9d4b384a66d", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Cutterhead and Cutting Tools Configurations in Coarse Grain Soils; DOI: 10.2174/1874836801711010182; license: cc-by."}, {"role": "assistant", "content": "Zhu, \"The design and construction of Komatsu combine shield used in sand-pebble stratum\", Constr. Mechanization, vol. 32, no. 11, pp. 67-69, 2011. [In Chinese]. [http://dx.doi.org/10.3969/j.issn.1001-1366.2011.11.025] [32] F. He, \"Analysis on adaptability of slurry shield cutterhead and cutters in Beijing railway underground connecting line project\", Tunnel Constr., vol. 31, no. 4, pp. 416-425, 2011. [33] W. Burger, \"Interventions and chamber access in pressurized face TBMs\", In: Rapid Excavation and Tunneling Conference Proceedings, 2011, pp. 1036-1047. [34] W. Burger, \"Hard rock cutterhead design\", In: Proceedings and Monographs in Engineering, Water and Earth Science, 2006, pp. 257-263. [35] F. Enrique, L.M. Juan, and S. Alejandro, \"Technical approach on big preparation to succeed on the Alaskan way project\", In: Rapid Excavation and Tunneling Conference Proceedings, 2011, pp. 102-118. © 2017 Li et al. This is an open access article distributed under the terms of the Creative Commons Attribution 4.0 International Public License (CC-BY 4.0), a copy of which is available at: https://creativecommons.org/licenses/by/4.0/legalcode. This license permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. View publication stats"}], "source_title": "Cutterhead and Cutting Tools Configurations in Coarse Grain Soils", "source_doi": "10.2174/1874836801711010182", "source_url": "https://openconstructionbuildingtechnologyjournal.com/VOLUME/11/PAGE/182/PDF/", "source_license": "cc-by", "zotero_item_key": "7KDIA2X2"}
{"id": "9751593d507f2597", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Detection of Ionospheric Scintillation Based on XGBoost Model Improved by SMOTE-ENN Technique; DOI: 10.3390/rs13132577; license: cc-by-sa."}, {"role": "assistant", "content": "2020, 125, 2019JA027487. 3. Karan, D.K.; Daniell, R.E.; England, S.L.; Martinis, C.R.; Eastes, R.W.; Burns, A.G.; Mcclintock, W.E. First zonal drift velocity measurement of Equatorial Plasma Bubbles (EPBs) from a geostationary orbit using GOLD data. J. Geophys. Res. Space Phys. 2020, 125, e2020JA028173. [CrossRef] 4. Martinis, C.; Daniell, R.; Eastes, R.; Norrell, J.; Smith, J.; Klenzing, J.; Solomon, S.; Burns, A. Longitudinal variation of post-sunset plasma depletions from the Global-scale Observations of the Limb and Disk (GOLD) mission. J. Geophys. Res. Space Phys. 2020, 126, 2020JA028510. 5. Liu, J.; Wang, W.; Burns, A.; Solomon, S.C.; Zhang, S.; Zhang, Y.; Huang, C. Relative importance of horizontal and vertical transports to the formation of ionospheric storm-enhanced density and polar tongue of ionization. J. Geophys. Res. Space Phys. 2016, 121, 8121–8133. [CrossRef] 6. Basu, S.; MacKenzie, E.; Basu, S. Ionospheric constraints on VHF/UHF communications links during solar maximum and minimum periods. Radio Sci. 1988, 23, 363–378. [CrossRef] 7. Sreeja, V.; Aquino, M.; Elmas, Z.G. Impact of ionospheric scintillation on GNSS receiver tracking performance over Latin America: Introducing the concept of tracking jitter variance maps. Space Weather 2011, 9, S10002. [CrossRef] 8. Aarons, J.; Whitney, H.E.; Allen, R.S. Global morphology of ionospheric scintillations. Proc. IEEE 1971, 59, 159–172. [CrossRef] 9. Knepp, D.L. Radar measurement of ionospheric scintillation in the polar region. Radio Sci. 2015, 50, 968–982. [CrossRef] 10. Jiao, Y.; Morton, Y.T. Comparison of the effect of high-latitude and equatorial ionospheric scintillation on GPS signals during the maximum of solar cycle 24. Radio Sci. 2015, 50, 886–903. [CrossRef] 11. Prasad, S.N.V.S.; Rama Rao, P.V.S.; Prasad, D.S.V.V.D.; Venkatesh, K.; Niranjan, K. Morphological studies on ionospheric VHF scintillations over an Indian low latitude station during a solar cycle period (2001–2010). Adv. Space Res. 2012, 50, 56–69. [CrossRef] 12. Banville, S.; Langley, R.B.; Saito, S.; Yoshihara, T. Handling cycle slips in GPS data during ionospheric plasma bubble events. Radio Sci. 2010, 45, 1–14. [CrossRef] 13. Ji, S.; Chen, W.; Weng, D.; Wang, Z.; Ding, X. A study on cycle slip detection and correction in case of ionospheric scintillation. Adv. Space Res. 2013, 51, 742–753. [CrossRef] 14. Taylor, S.; Morton, Y.; Jiao, Y.; Triplett, J.; Pelgrum, W. An improved ionosphere scintillation event detection and automatic trigger for a GNSS data collection system. In Proceedings of the Institute of Navigation International Technical Meeting 2012, Newport Beach, CA, USA, 30 January–1 February 2012; pp. 1563–1569. 15. Dubey, S.; Wahi, R.; Gwal, A.K. Ionospheric effects on GPS positioning. Adv. Space Res. 2006, 38, 2478–2484. [CrossRef] 16. Linty, N.; Farasin, A.; Favenza, A.; Dovis, F. Detection of GNSS ionospheric scintillations based on machine learning decision tree. IEEE Trans. Aerosp. Electron. Syst. 2019, 55, 303–317. [CrossRef]\n[page 21] Remote Sens. 2021, 13, 2577 21 of 22 17. Mushini, S.C.; Jayachandran, P.T.; Langley, R.B.; MacDougall, J.W.; Pokhotelov, D. Improved amplitude- and phase-scintillation indices derived from wavelet detrended high-latitude GPS data. GPS Solut. 2012, 16, 363–373. [CrossRef] 18. Ouassou, M.; Kristiansen, O.; Gjevestad, J.G.O.; Jacobsen, K.S.; Andalsvik, Y.L. Estimation of scintillation indices: A novel approach based on local kernel regression methods. Int. J. Navig. Obs. 2016. [CrossRef] 19. Jiao, Y.; Hall, J.; Morton, Y. Performance evaluations of an equatorial GPS amplitude scintillation detector using a machine learning algorithm. In Proceedings of the 29th International Technical Meeting of the Satellite Division of the Institute of Navigation, Portland, OR, USA, 12–16 September 2016; pp. 195–199. 20. Jiao, Y.; Hall, J.J.; Morton, Y.T. Automatic equatorial GPS amplitude scintillation detection using a machine learning algorithm. IEEE Trans. Aerosp. Electron. Syst. 2017, 53, 405–418. [CrossRef] 21. Jiao, Y.; Hall, J.; Morton, Y. Automatic GPS phase scintillation detector using a machine learning algorithm. In Proceedings of the 2017 International Technical Meeting of The Institute of Navigation, Monterey, CA, USA, 30 January–2 February 2017; pp. 1160–1172. 22. Jiao, Y.; Hall, J.J.; Morton, Y.T. Performance Evaluation of an automatic GPS ionospheric phase scintillation detector using a machine-learning algorithm. Navig. J. Inst. Navig."}], "source_title": "Detection of Ionospheric Scintillation Based on XGBoost Model Improved by SMOTE-ENN Technique", "source_doi": "10.3390/rs13132577", "source_url": "https://www.mdpi.com/2072-4292/13/13/2577/pdf?version=1625212202", "source_license": "cc-by-sa", "zotero_item_key": "4YEIWMK3"}
{"id": "9d736f0a7e11db7c", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Two-phase multi-model automatic brain tumour diagnosis system from magnetic resonance images using convolutional neural networks; DOI: 10.1186/s13640-018-0332-4; license: cc-by-sa."}, {"role": "assistant", "content": "Ardeshir, in 2003 International Conference on Visual Information Engineering, VIE 2003. Boosted ECOC ensembles for face recognition (IET, Guildford, 2003), pp. 165–168 41. J. Kittler, R. Ghaderi, T. Windeatt, J. Matas, in Proceedings of the 2001 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, CVPR 2001, vol 1. Face verification using error correcting output codes, (2001), pp. 755–760 42. R. Ghani, in Proceedings of the 2001 IEEE International Conference on Data Mining. Combining labeled and unlabeled data for text classification with a large number of categories, (2001), pp. 597–598 43. J. Zhou, C. Y. Suen, in Eighth International Conference on Document Analysis and Recognition, ICDAR’05, vol 1. Unconstrained numeral pair recognition using enhanced error correcting output coding: a holistic approach, (2005), pp. 484–488 44. F. Deng, S. Guo, R. Zhou, J. Chen, Sensor multifault diagnosis with improved support vector machines. IEEE Trans. Autom. Sci. Eng. PP(99), 1–11 (2015) 45. R. Girshick, Fast R-CNN. International Conference on Computer Vision (ICCV), Santiago, Chile, 11–18 December, 2015. (IEEE, Santiago, 2015), pp. 1440–1448 46. P. Reverdy, N. E. Leonard, Parameter estimation in softmax decision-making models with linear objective functions. IEEE Trans. Autom. Sci. Eng. 13(1), 54–67 (2016) 47. C. M. Bishop, Pattern recognition and machine learning. (Springer, 2006) 48. B. Yuan, in 2016 29th IEEE International System-on-Chip Conference (SOCC). Efficient hardware architecture of softmax layer in deep neural network (IEEE, Seattle, 2016), pp. 323–326"}], "source_title": "Two-phase multi-model automatic brain tumour diagnosis system from magnetic resonance images using convolutional neural networks", "source_doi": "10.1186/s13640-018-0332-4", "source_url": "https://jivp-eurasipjournals.springeropen.com/track/pdf/10.1186/s13640-018-0332-4", "source_license": "cc-by-sa", "zotero_item_key": "2ZZSU54R"}
{"id": "b158bda876b09ce8", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Automated Excavator Based on Reinforcement Learning and Multibody System Dynamics; DOI: 10.1109/access.2020.3040246; license: cc-by-sa."}, {"role": "assistant", "content": "The rbn vector has the following form: rbn = \u0002 x y z e0 e1 e2 e3 \u0003T (22) where x, y and z are XYZ-axes positions of the body, e0, e1, e2 and e3 are the Euler parameters of the body. In rbn vectors stored locations and rotations of the Upper Carriage, Lift Boom, Dipper Arm and Bucket bodies. The rg stores the positions of the goals and have a size of 6. It has the following form: rg = \u0002 xg1 yg1 zg1 xg2 yg2 zg2 \u0003T (23) where xg1, yg1, zg1 are positions of Goal To Soil and xg2, yg2, zg2 are positions of Goal To Hopper markers. In total, the observation vector consists of 42 parameters. Before development of the reward function, it is important to understand the task given to the agent. Multiple excavator operations exist in real life. For instance, in construction of underground pipelines, the excavator must create a trench, which requires so-called bucket leveling. On the other hand, mining work requires movement of soil from the ground into a truck without hitting the truck. Clearly, each operation requires a different approach, but they all involve movement of soil from one point to another. Therefore, movement of soil was chosen as the basis of the experiment. The selected task consists of three steps: loading the bucket near the loading point, moving to the unloading point, and unloading the bucket. In the first step, the agent moves the bucket to the loading position and grabs the soil. During this step, it is important to move the bucket quickly and to grab the maximum amount of soil. After loading of the bucket, the soil must be moved to the unloading point over the hopper. This step requires the agent to move the bucket with little or no loss of soil. In the final step, the agent must unload the maximum amount of soil without hitting the hopper. The setup of the simulation is shown in Fig.6. The environ- ment consists of the soil model, points of interest for the task, VOLUME 8, 2020 214003\n[page 7] I. Kurinov et al.: Automated Excavator Based on RL and Multibody System Dynamics FIGURE 6. Simulation setup. The excavator is located uphill of the hopper and the unloading hopper is positioned to the right of the excavator. and a hopper. The soil model allows collection and transfer of soil, which is essential information for training. Points of interest provide the agent with information about soil grabbing and unloading positions. The hopper is equipped with two sensors. The first sensor collects data about the mass in the hopper, and the second sensor tracks the number of collisions with the bucket. Using this setup, it is possible to track all the variables needed for formulation of a reward function. The reward function of the excavator model utilizes three components. The first component is a reward for reaching the point where the bucket should be loaded. This point is marked as a blue dot in Fig.6. The second component is a reward for the mass collected in the bucket during loading. For preventing abuse of soil mass reward, it is inverse propor- tional to episode time. The third component utilizes the mass in the hopper, which is penalized by the number of collisions between the hopper and bucket. Thus, the reward function can be written as: r (t) = e−l + kmbntn + 1 nc mhn (24) where r(t) is a reward at the time t, k is the mass reward coef- ficient, l is the distance to the current point of interest, mbn is the normalized mass in the bucket, mhn is the normalized mass in the hopper, and nc is the number of collisions with the hopper. IV. RESULTS The excavator model is highly nonlinear and includes mechanics, hydraulics, sensor measurements, a chang- ing environment and process visualization. Consequently, the training process is challenging. In the simulation, the agent learnt acceptable performance after 105 steps, which took around two days of continuous training. Comput- ing efficiency was affected by the deformable ground sub- model, which could not be updated without reloading the simulation. Therefore, a significant amount of training time was spent on reloading the simulation file. FIGURE 7. Result of the training process. Fig.7 shows the result of the training process. As can be seen from the figure, the excavator grabs soil at the desired loading position."}], "source_title": "Automated Excavator Based on Reinforcement Learning and Multibody System Dynamics", "source_doi": "10.1109/access.2020.3040246", "source_url": "https://ieeexplore.ieee.org/ielx7/6287639/6514899/09268069.pdf", "source_license": "cc-by-sa", "zotero_item_key": "7F2IQNFD"}
{"id": "1f655e264686d367", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Spectral Angle Mapping and AI Methods Applied in Automatic Identification of Placer Deposit Magnetite Using Multispectral Camera Mounted on UAV; DOI: 10.3390/min12020268; license: cc-by-sa."}, {"role": "assistant", "content": "Thereafter, the user pinpoints the area of significance, which in this case was any area within the multispectral image scene known to be occupied by magnetite iron sand. Lastly, the algorithm performs a cosine similarity calculation for every pixel in each image scene to determine whether or not there are similarities between the reference and subsequent subjects [11]. To determine the cosine similarity threshold, Equation (1) is used. a = cos−1 \u0012 t × t0 ∥t∥× ∥t0∥ \u0013 (1) where a is the threshold for variables related to the reference, t0 is the reference spectrum, and t is the subject’s true spectrum. Having successfully stacked the 6 bands TIFF and JPG qualitative images, as well as performing SAM analysis, the images are then converted to csv format. This data conver- sion provides quantitative values pertaining to magnetite iron sand, and non-magnetite beach sand spectral characteristics, across the 6 multispectral bands. This data structure is useful in training AI algorithms. 2.4. Benefits of Employing AI Methods in Magnetite Identification In order to determine regions of an image where magnetite iron sands exist post SAM analysis, ML and DL models are an upheld method by which this task can be performed quantitatively [15]. This step involves pre-processing the data before training the AI algorithms. Pre-processing is important as it cleans the data and makes it easier for the algorithms to interpret it; for this study, the pre-processing step involved the extraction of the SAM segmentation maps per image. This simply means magnetite and non-magnetite post SAM spectral variables, given their cosine similarity thresholds, are assigned labels so as to take advantage of supervised AI algorithms. Hence, this process is referred to as image segmentation, and is simply defined as the labelling of each pixel in an image scene [15]. Pixels within and outside the threshold are assigned magnetite and non-magnetite labels, respectively. Since SAM is mostly a visual interpretation, ground truthing is needed in order to employ AI algorithms. For this reason, the SAM segmentation map is extracted together with two ground truths for magnetite and non-magnetite pixels. Ground truths are image pixels known to be 100% occupied by magnetite and non-magnetite pixels per image. Since each image is 6 bands deep, the result is a total of 12 ground truths. Ground truth for magnetite is defined as the target, whereas non-magnetite ground truth is defined as noise. Labelled datasets guide and allow algorithms to take advantage of both the global and local structures of the data, which in turn results in quicker and more direct classifica- tions [10]. This in essence means that for each image, the csv file contains labelled pixels for both magnetite iron sands (together with target ground truth), as well as non-magnetite beach sand (together with noise ground truth). Using this data as input, the algorithms are subsequently trained and outputs pertaining to the classification capabilities of the models in deducing magnetite iron sands and non-magnetite beach sand are presented. 2.4.1. How Machine Learning Algorithms Operate DL is a subset of ML in the broader subject of AI [10]. Though these algorithms are useful in multitudes of data-related tasks such as the classification problem being investigated in this study, they differ in several ways. In order to infer the top performing model at each UAV flight height, both ML and DL AI algorithms are trained. ML algorithms are mathematical engines of AI, which means these algorithms attribute their classification abilities to perceived mathematical relationships present within the data [19]. In simple terms, ML algorithms try to fit the data within a particular pattern,\n[page 7] Minerals 2022, 12, 268 7 of 19 which can be described using mathematical functions. Since these algorithms are relatively easy to employ, several ML algorithms were trained with regard to the ground truths for target and noise pixel variables. These algorithms were Decision Trees (often abbreviated as Trees) and Ensemble algorithms. Decision Tree algorithms were employed as they are non-parametric supervised learning methods, used for classification where the goal is to make predictions by learning simple piecewise constant approximation decision rules from data features [25]. Therefore, deeper decision rules are referred to as fine, whereas shallow and less complex decision rules are referred to as course."}], "source_title": "Spectral Angle Mapping and AI Methods Applied in Automatic Identification of Placer Deposit Magnetite Using Multispectral Camera Mounted on UAV", "source_doi": "10.3390/min12020268", "source_url": "https://www.mdpi.com/2075-163X/12/2/268/pdf?version=1645347690", "source_license": "cc-by-sa", "zotero_item_key": "VPIE7856"}
{"id": "a110e54f0bb84a03", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Reinforcement learning based process optimization and strategy development in conventional tunneling; DOI: 10.1016/j.autcon.2021.103701; license: cc-by."}, {"role": "assistant", "content": "Ground type Specific weight [kN/m3] Cohesion [kPa] Friction angle [◦] Permeability [m/s] gt1 24 23 20 10−5 gt2 25 40 30 10−5 G.H. Erharter et al.\n[page 5] Automation in Construction 127 (2021) 103701 5 The output of the third convolutional layer is flattened/vectorized before it is fed into the dense layer and we used the mean squared error as a loss function. 3.2.1. Actions The agent can choose from 8 different actions (Table 3). Possible actions are either top heading – or bench excavation with advance lengths of either 2 or 4 m. In each step the simulation carries out the following operations: ▪ the agent chooses and executes one of the 8 actions of Table 3 and the respective length of the generated geological section is revealed ▪ it is calculated/checked if the new position of the excavation face is within a stable area or not (details in Section 3.3.2) As given in Section 2.1, the support ahead of the face has the effect on the simulation that, there cannot be unstable conditions as long as the excavation face is within the supported area. Because of technical rea­ sons - mostly related to the reward system (see Table 4) - each action is assigned an “action code” (a). An a < 200 denotes top heading exca­ vation and a ≥200 bench excavation. The choice of actions of Table 3 is based on practical engineering experience on the one hand and technical limitations of the RL agent on the other hand. On the practical side, especially the advance length is Fig. 4. Top row: an exemplary unique geological section, where brown indicates weak (gt1) and blue stronger rock (gt2). The positions of the top heading and bench are at 165.0 m and 125.0 m respectively. Bottom row: the random walk that is used to generate the geological section. Values above 0.5 are converted to gt2 and below to gt1. Note that the x-axis is the tunnel length in decimeters which corresponds to the number of datapoints of the random walk. Table 2 Used hyperparameters for the DQN agent. Except for the exploration decay (see Section 3.2.1) all are identical in their meaning as the extended data Table 1 in [14]. Hyperparameter Value Replay memory size 100,000 Replay start size 1000 Minibatch size 64 Discount (γ) 0.99 Target network update frequency 10 Initial exploration (ε) 1 Final exploration (ε) 0.05 Learning rate 0.00025 Gradient momentum 0.95 Exploration decay (εd) 0.99997 Fig. 5. Schematic representation of the DQN agent’s ANN architecture. Note the visualization of rockmass-matrix and the support-matrix to the left. The numbers below each layer are the respective shape of the layer’s weights. Dashed connection lines between layers are only for illustrational purposes. Symbols at the output layer represent the eight possible actions (ordered as in Table 3) that are chosen via Q-values by the agent. Table 3 The eight possible actions the agent can choose from. Action code (a) Excavation Advance length [m] Face support 110 top heading 2 no 112 top heading 2 yes 150 top heading 4 no 152 top heading 4 yes 200 bench 2 no 202 bench 2 yes 220 bench 4 no 222 bench 4 yes G.H. Erharter et al.\n[page 6] Automation in Construction 127 (2021) 103701 6 highly influential on the stability of the excavation with longer advance lengths being more prone to failures than short ones (e.g. advance lengths in the Austrian standard ¨ONORM B2203–1 [46]: 1.0, 1.3, 1.7, 2.2, 3.0 and 4.0 m). After experimenting with advances lengths based on the standards in the beginning, we experienced that a big number of actions leads to a deterioration of the performance/confuses the agent in this model setup. This is also in accordance with other studies which have found that special measures/adaptations of the agent are necessary if the action space becomes increasingly complex [47,48]. The final set of actions in Table 3 therefore aims at giving the agent realistic options to choose from while also keeping the number of actions small (see the outlook in Section 7 for a discussion on increasing the number of actions). During an episode, the actions are chosen based on an “ε-greedy action selection process”, (see also “exploration vs. exploitation trade­ off”, e.g. [26]). ε is the exploration rate and is initially set to 1 (see “initial exploration in Table 2)."}], "source_title": "Reinforcement learning based process optimization and strategy development in conventional tunneling", "source_doi": "10.1016/j.autcon.2021.103701", "source_url": "https://www.sciencedirect.com/science/article/pii/S0926580521001527", "source_license": "cc-by", "zotero_item_key": "RXHUVV3X"}
{"id": "96648044d41876ca", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Application of machine learning in predicting the rate-dependent compressive strength of rocks; DOI: 10.1016/j.jrmge.2022.01.008; license: cc-by-sa."}, {"role": "assistant", "content": "The procedure for the machine learning prediction of compressive strength followed that in Fig. 2. We quantitatively assessed the sensitivity of the strength prediction to each excluded parameter based on the percentage increments of four performance metrics (P1j, P2j, P3j, and P4j): P1j ¼ MAEj \u0001 MAE0 MAE0 \u0005 100% (18) P2j ¼ MXAEj \u0001 MXAE0 MXAE0 \u0005 100% (19) P3j ¼ MAPEj \u0001 MAPE0 MAPE0 \u0005 100% (20) P4j ¼ RMSEj \u0001 RMSE0 RMSE0 \u0005 100% (21) where the subscripts j and 0 refer to the performance metrics ob- tained from the models with five and six input parameters, respec- tively; and j(1, 2,. and6)corresponds to staticcompressive strength, Fig. 6. Comparison between the measured and predicted strengths from (a) SVM- based, (b) BPNN-based, and (c) RF-based models. The histograms on the upper and right axes show the density distributions of strength values. M. Wei et al. / Journal of Rock Mechanics and Geotechnical Engineering 14 (2022) 1356e1365 1361\n[page 7] P-wave velocity, specimen dimension, grain size, bulk density, and strain rate, respectively. For example, P26 denotes the percentage incrementof the MXAEvaluewhen the strain rate isexcludedinthese models. If P26 > P25, the strain rate plays a more critical role in pre- dicting the compressive strength than the bulk density. The percentage increments of the performance metrics plotted in Fig. 8 exhibit that the strength prediction result is more sensitive to the strain rate than the other input parameters. When the strain rate is excluded from the input parameters of the machine learning models, the performance metrics except the MXAE show larger percentage increments, compared with those excluding the other five parameters. The exception of the MXAE is most likely attrib- uted to one or a few data points with unusual values. The strain rate controlled by the loading condition is less dependent on the geometrical, physical, and mechanical properties of rocks and should be included in the strength prediction. This observation is similar to that observed from the Pearson correlation coefficient. The static compressive strength is correlated with P-wave velocity, grain size, and bulk density (Fig. 1). If the static compressive strength is excluded, the impact of the static compressive strength on the strength prediction can be partially reflected by P-wave velocity and bulk density. That is probably why the performance metrics P11, P21, P31, and P41 approach zero. The results indicate that the influence of specimen dimension is relatively minor, which may be addressed insufficiently, particularly the specimen diameter, as the bar diameter in the SHPB setup is mostly limited to a range of 25e100 mm. Also, the aspect ratio as a related factor is not well studied due to the requirement of dynamic stress equilibrium in rock specimens with appropriate aspect ratios in the SHPB tests. 4.2. Rate dependence of rock strength Our results demonstrated that the RF-based model shows a better performance than the SVM- and the BPNN-based models (Fig. 7), and the exclusion of the static compressive strength has a negligible impact on the prediction accuracy (Fig. 8). Thus the RF- based model with five input parameters was adopted, except the static compressive strength, to further discuss the rate depen- dence of rock strength following the procedure in Fig. 2. This model considers a series of fine-grained sandstone specimens (Yu et al., 2021) with the same aspect ratio and three different specimen diameters (12.5 mm, 25 mm, and 50 mm). The pre- diction result shows that the compressive strength is relatively constant under the strain rate lower than 20 s\u00011 and drastically increases under the strain rate higher than 20 s\u00011 (Fig. 9a). The increase in compressive strength is mainly attributed to the transition from inter-crystalline fracture to intra-crystalline frac- ture, the amplified energy absorption, and the inertial effect (Zhang et al., 2000). The trend of strength increase is similar for different specimen diameters, and the compressive strength de- creases with increasing specimen diameter. Then, we used the dynamic increase factor (DIF), defined as the ratio of the dynamic to static strength, to highlight the effect of specimen diameter on the compressive strength. The three curves start to fluctuate when the strain rate exceeds 20 s\u00011 (Fig. 9b). From 20 s\u00011 to 50 s\u00011, the DIF values of these specimens with the three diameters increase in a similar trend."}], "source_title": "Application of machine learning in predicting the rate-dependent compressive strength of rocks", "source_doi": "10.1016/j.jrmge.2022.01.008", "source_url": "https://www.sciencedirect.com/science/article/pii/S167477552200049X", "source_license": "cc-by-sa", "zotero_item_key": "KNCWQQJD"}
{"id": "b6725866f5be6a2a", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Two-phase multi-model automatic brain tumour diagnosis system from magnetic resonance images using convolutional neural networks; DOI: 10.1186/s13640-018-0332-4; license: cc-by-sa."}, {"role": "assistant", "content": "In the pre- processing stage, we detect the tumour region from the segmented image using a bounding box method. The T2 image with the tumour bounding box are provided to the R-CNN block to train the network. As a post-processing step, we measure the properties of regions in the tested image and select the region most intersected with the R-CNN as the tumour region. 3.2.1 5-layer region-based convolutional neural network (R-CNN) R-CNNs offer high object detection accuracy. The R- CNN method trains a CNN to classify the image regions into an object or a background category. It uses the log loss approach to fine-tune the image regions into objects. Then, an SVM approach is applied to detect objects [45]. Table 4 Comparison of the performance of the three CNN models used in the tumour detection phase Network model Layer number (type) Accuracy (%) AlexNet Layer 9 (pool) 99.55 VGG-16 Layer 32 (pool) 92.95 VGG-19 Layer 38 (pool) 95.15\n[page 7] Abd-Ellah et al. EURASIP Journal on Image and Video Processing (2018) 2018:97 Page 7 of 10 Table 5 Comparison of the obtained tumour detection accuracy from the proposed CNN-based system to previous work in which a classical approach was used with the same database as a unified benchmark Detection system Performance evaluation metrics (%) Sensitivity Specificity Accuracy BA PPV NPV Abd-Ellah et al. [18] 83.43 25.00 66.96 54.21 73.91 37.20 Proposed method 99.38 100.0 99.55 99.69 100 98.46 The R-CNN architecture used in the tumour localiza- tion phase consists of five layers. The network first pro- cesses the whole image using the image-input layer (layer 1) and then converts the input data with a size of 28×28×1 into a convolutional layer (layer 2) with a filter size of 5 and filter number of 20 for detecting multiple class-specific objects. A fully connected layer (layer 3) with two outputs is used to predict a single-box for tumour localization. The softmax layer (layer 4) is also known as the normal- ized exponential. The output unit activation function is the softmax function [46–48], as presented in Eq. 6: yr = exp(ar) \u0003k j=1 exp(aj) (6) where a is the input vector with k dimensions and y is the output vector with k dimensions. The denomina- tor \u0003k j=1 exp(aj) normalizes the output to ensure that \u0003k r=1 yr = 1 and 0 ≤yr ≤1, which can be represented graphically as a layer with k neurons. The classification layer used as the final layer is named after the loss used for the training process. It returns a neural network classification output. For multi-class clas- sification, the loss (error) function is the cross entropy function shown in Eq. 7: E(θ) = − n \u0002 i=1 k \u0002 j=1 tij ln yj(xi, θ) (7) where θ is the parameter vector, tij is the indicator that the i sample belongs to the j class, and yj(xi, θ) is the output for sample i. Table 6 Comparison of the proposed CNN-based brain tumour detection approach against some former state-of-the-art methods CAD system Total images Classification method Accuracy (%) Da et al. [26] 10 DNN 83.0 Gao et al. [27] 285 2D and 3D CNNs 87.6 Yan Xu et al. [28] 45 CNNs 97.50 Proposed method 349 ECOC-SVM 99.55 4 Results and discussion 4.1 Experimental environment setup The image database used for evaluating the tumour detec- tion phase was extracted from the RIDER Neuro MRI database [17]. The input database consists of brain MRIs from 19 patients. These images were obtained approxi- mately 2 days apart. T1-weighted and T2-weighted MRIs were used. The database contains a total of 349 MRIs, including 109 normal images and 240 abnormal images. A set of 43264 features was extracted from each MRI image which later was used to train the SVM. The SVM was trained using 122×43264 features in total. The specifica- tions of the utilized database are presented in Table 2. A sample of the MRIs database used in the experimental work is shown in Fig. 3. To evaluate the tumour localization phase, a dataset of 804 3D MRIs extracted from the BraTS 2013 database was used for the localization accuracy assessment. The database consists of 20 high-grade (HG) and 10 low-grade (LG) patients and T1, T1c, T2, and FLAIR type MRI modalities. A set of 593 MRIs was used for R-CNN train- ing and validation, and a set of 211 MRIs was used for the testing dataset. The 3D images were converted into 2D slices, and the T2 MRI modality was used."}], "source_title": "Two-phase multi-model automatic brain tumour diagnosis system from magnetic resonance images using convolutional neural networks", "source_doi": "10.1186/s13640-018-0332-4", "source_url": "https://jivp-eurasipjournals.springeropen.com/track/pdf/10.1186/s13640-018-0332-4", "source_license": "cc-by-sa", "zotero_item_key": "2ZZSU54R"}
{"id": "fe6cca52c0de84e1", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Cutterhead and Cutting Tools Configurations in Coarse Grain Soils; DOI: 10.2174/1874836801711010182; license: cc-by."}, {"role": "assistant", "content": "The average wear coefficients for scrapers of the spoke type and the spoke- plate type cutterheads are close and are 0.0081 and 0.0068 mm/km respectively. The wear coefficients for scrapers in the two cases are both less than 0.01 mm/km, only 1/4 of those for precut bits, as shows that the scrapers play the role of ripping off, agitating and not shaving off ground. So, it is concluded that the two opening ratios of 39% and 75.5% have almost the same influences on tool wears and are both justifiable. However, the opening ratio of 39% is more reasonable if more considerations are given to the cutting face stability and shield machine manipulation. If shield tunneling will be performed in the coarse grain soils rich in groundwater and maintaining cutting face stability is a demanding job, a spoke-plate type, or plate type cutterhead with high opening ratio, such as 39%, for the specific ground condition is strongly recommended. As to deciding the site-specific opening ratio of a cutterhead, a good estimation of the size of grains in soils takes the lead. In Chengdu, most metro tunnels are constructed in water-rich coarse grain ground and the plate type cutterheads are usually employed mainly for the sake of more reliably maintaining stability of the cutting face. 4.2. Configuration of Cutting Tools It is observed from the previous construction cases that there are two typical types of configurations of tools for the cutterheads used in coarse grain soils; one is for the spoke type or spoke-plate cutterhead with high opening ratio and the other is for the plate type cutterhead with an opening ratio less than 30%. Dependent on specific characteristics of the ground conditions and the suitable ground excavation mechanism, the cutting tools, including precut bits, shell bits, scrapers and a center fish-tail bit, are versatile combinations of tools for spoke type and spoke-plate cutterheads with high opening ratio in Beijing and Shenyang coarse grain soils, which has been displayed by the relevant construction cases. With the help of loosening soils by precut bits, shell bits and the center fish-tail bit beforehand, the oversized particles (cobbles and boulders) can be easily ripped out of the soil matrix causing less thrust and less torque of a shield machine as well as rapid tunnel construction. Of course, a shaftless ribbon type screw conveyor is often indispensable to smooth discharge of the large sized grains in the excavated materials. The plate type cutterheads with opening ratio about 30% are widely used in Chengdu metro tunnel construction mainly on account of the soils rich in groundwater. The tools used at the beginning of tunnel construction were disc cutters in combination with scrapers presented in (Fig. 6a). Standing out a little farther in profile, the installed disc cutters were to break big cobble/boulders in soils into pieces and loosen ground in front of the cutterhead at the same time, but often suffered severe wears only after a driving less than 100 m and frequent changing tools work had to be performed. It was one of the most important improvement measures taken in Lot 22, Lot 11 of Line 2 and Lot 4 of Line 1 of Chengdu metro that precut bits were added or used to replace some disc cutters. The driving length was increased to 300 m or even more before necessary interventions of changing tools. Therefore, for the plate type cutterheads, the combination of precut bits and scrapers with/without disc cutters may be a better choice. Because scrapers are vulnerable to big cobbles and boulders in ground excavation, proper height differences between precut bits and scrapers are especially important to play different roles of the two types of tools. For example, the differences are in the ranges of 40-50 mm in Beijing sandy gravel/cobble soils. Meantime, in view of good performances of the previous cutterheads with precut bits, the precut bits should be strong enough to help reduce wears of the scrapers to the minimum. 4.3. Use of Single Disc Cutters with Inserts It is very common for shield tunneling in the coarse grain soils that the installed center disc cutters and single disc cutters suffer excessive abnormal wears. At a low rotation speed of the cutterhead, the center disc cutters are easily\n[page 16] 196 The Open Construction and Building Technology Journal, 2017, Volume 11 Li et al. jammed by clogging of the cutterhead (see Fig. (20))."}], "source_title": "Cutterhead and Cutting Tools Configurations in Coarse Grain Soils", "source_doi": "10.2174/1874836801711010182", "source_url": "https://openconstructionbuildingtechnologyjournal.com/VOLUME/11/PAGE/182/PDF/", "source_license": "cc-by", "zotero_item_key": "7KDIA2X2"}
{"id": "88de11bddd560337", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Detection of Ionospheric Scintillation Based on XGBoost Model Improved by SMOTE-ENN Technique; DOI: 10.3390/rs13132577; license: cc-by-sa."}, {"role": "assistant", "content": "Based on this research, Franzese et al. (2020) [29] proposed semi-supervised scintillation detection using the DeepInformax technique, which presents a validation accuracy in accord with that of the decision tree model. The decision tree and random forest models are vulnerable to the problem of overfitting if the high model complexity is designed improperly. More importantly, these methods mentioned above rarely consider the imbalanced phenomenon of different intensities of scintillation events [16,19–25,29].\n[page 3] Remote Sens. 2021, 13, 2577 3 of 22 Insufficient data related to strong scintillation may cause great trouble in model training. Moreover, the events to be tested with different intensities may show different degrees of imbalance. Such imbalance in different intensities of scintillation events is rarely taken into consideration. Among the occasional ionospheric scintillation events, strong scintillation is a minority class compared to weak and moderate scintillation. For this class of data, few classification algorithms can accurately describe the inherent characteristics due to the lack of information about the minority class, which causes the decision boundary to be greatly compressed in the classification model [30]. Although the overall accuracy of the final model is relatively high, there is some missed detection for strong scintillation events. It can be optimized by solving the imbalance of the dataset to reduce the problem of missing detection points for strong scintillation events in the testing set. In this paper, an improved machine learning method is proposed for improving the automatic detection accuracy for strong ionospheric scintillation, and it is compared in detail with the present decision tree and random forest, which show high accuracy for the validation set [16]. There is also a brief comparison with an SVM and CNN, which have recently been utilized in relevant research [19–28]. The proposed approach can provide significant guidance for designing robust GNSS receivers as well as research on atmospheric layers and space weather. Several aspects of this approach can be described as follows: (1) To propose a high-performance detection method for XGBoost based on the deci- sion tree algorithm, assuring good overall detection accuracy for three intensities of ionospheric scintillation; (2) To compare with decision tree and random forest based on cross-validation, proving the superior accuracy of the XGBoost algorithm and mitigation of the overfitting problem; (3) To compare different resampling techniques for an imbalanced dataset consisting of a majority of weak/medium scintillation events and a minority of strong scintillation events, proving the great performance of SMOTE-ENN according to evaluation indicators; (4) To make a brief overall comparison with the decision tree, random forest, SVM and CNN, evaluating the performance in terms of accuracy, computational load and applicability, focusing on detailed comparison with the first two methods with high performance hereafter; (5) To evaluate the performance of the proposed improved method on different de- grees of imbalanced training datasets and testing datasets, respectively, proving the effectiveness in enhancing the detection accuracy for strong scintillation events. In this work, a strategy of using data processing technology for imbalanced iono- spheric scintillation events is proposed, and we applied our strategy to space weather detection for the first time, which successfully improved the detection accuracy for strong scintillation events. Additionally, an improved machine learning algorithm with proven high overall accuracy for scintillation events of different intensities is proposed. The general overviews are described in Section 2, including an introduction to ionospheric scintillation, the data collection system and feature extraction. Section 3 presents an improved machine learning method, giving a theoretical introduction to the XGBoost algorithm and SMOTE- ENN technique. Section 4 validates the optimal detection performance of the proposed approaches, showing and evaluating the quantitative results obtained from the training and testing on datasets with different degrees of imbalance. Section 5 discusses various training and testing cases, and Section 6 draws the main conclusions. 2."}], "source_title": "Detection of Ionospheric Scintillation Based on XGBoost Model Improved by SMOTE-ENN Technique", "source_doi": "10.3390/rs13132577", "source_url": "https://www.mdpi.com/2072-4292/13/13/2577/pdf?version=1625212202", "source_license": "cc-by-sa", "zotero_item_key": "4YEIWMK3"}
{"id": "e1ca16cb377509ad", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Reinforcement learning based process optimization and strategy development in conventional tunneling; DOI: 10.1016/j.autcon.2021.103701; license: cc-by."}, {"role": "assistant", "content": "Before every move, a number (r) is drawn from a random uniform distribution in the range between 0 and 1. r governs the probability for a move to be a random action (if r ≤ε) or to be based on the agent’s prior experience (if r > ε). Throughout training ε will decay following Eq. (2), where the new ε for the next episode (εi+1) is computed by the previous ε (εi) times a constant – the epsilon decay (εd). εi+1 = εi*εd (2) This process guarantees that the agent can explore the environment in the beginning and shifts towards more exploitation of its knowledge towards the end of training. We set the minimum ε to 0.05 below which there is no more decay and ε will be kept constant (see “final explora­ tion” in Table 2). 3.3. Environment As the DQN agent is the RL pendant to the real life geotechnician, the environment is the RL representation of the rockmass and of logistical processes of the construction site (state) as well as the feedback that is received for better or worse excavation performance (reward). Based on the agent’s actions, its main tasks are to update and yield the current state of the construction-site and to provide feedback. 3.3.1. States The state that is observed by the DQN agent (see Section 3.2) is a hypermatrix of the shape 2 × 2100 × 2 which represents a geological section and a section that shows where support is already installed. The hypermatrix is structured in the following way (see Fig. 6): 2 rows for top heading and bench respectively; 2100 columns for the total length of the tunnel in decimeters plus an additional area beyond the break­ through (see below); 2 channels for the geological section and the sec­ tion with the installed support respectively. The values of the hypermatrix – originally ranging from 0 to 2 (see Section 3.1) - are then scaled between 0 and 1. In the channel of the geological section, 0 therefore translates to “not yet excavated area” and 0.5 and 1 represent gt1 and gt2 respectively (see Table 1 and Section 3.1). In the channel of the already installed support, 0 means “no installed support” and 1 means “installed support”. There are 2 terminal states in the simulation (i.e. states that lead to abortion of the simulation after their occurrence): ▪ A breakthrough is achieved if and only if both posth and posbi are ≥tl. In this case a breakthrough reward is given (see next section) and the simulation is finished. ▪The second terminal state is a timeout which is set to 200 ac­ tions; i.e. the agent must achieve a breakthrough within less than 200 actions or otherwise the simulation is aborted, and a negative breakthrough reward is given. The environment is designed in a way that once posth or posbi are ≥tl, the position of this part of the excavation is not updated anymore even if it is further excavated. We introduced the “timeout” as we observed in the experimental phase of the study that the agent sometimes reaches a breakthrough with the top heading or the bench but keeps on excavating in the already excavated part of the tunnel. On the one hand this causes the episode to be infinitely long, and on the other hand this leads to excessively large negative penalties which negatively affect the training process. We set the total length of the sections to 210 m (i.e. tunnellength + maximum possible length of face support) to give the agent the freedom to use actions that install face support even for the last few blasts. From a “real world tunneling” perspective this does not make sense and in the simulation, this would lead to additional penalties, however, the goal is that the agent learns things like this and is not forced to do so. 3.3.2. Reward system Rewards have the purpose to tell the agent if its actions are beneficial for the total reward. The ultimate goal of the agent is to maximize the return (i.e. sum of the rewards throughout the episode – here: excavate the whole tunnel). During an episode, a reward (also called penalty in case of a negative reward) is given for each individual move that the agent takes [26]. The action is chosen to maximize the expected return of discounted rewards. Reward systems in RL can vary greatly depending on the given task."}], "source_title": "Reinforcement learning based process optimization and strategy development in conventional tunneling", "source_doi": "10.1016/j.autcon.2021.103701", "source_url": "https://www.sciencedirect.com/science/article/pii/S0926580521001527", "source_license": "cc-by", "zotero_item_key": "N5Q33JID"}
{"id": "844ecad1a872fd7d", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Reinforcement learning based process optimization and strategy development in conventional tunneling; DOI: 10.1016/j.autcon.2021.103701; license: cc-by."}, {"role": "assistant", "content": "While [14], who trained their DQN to play classical Atari 2600 games, took a modified version of the original games’ scoring systems as a means to deal with the reward in each timestep, [15] only gave a reward of +1 or − 1 at the end of the game depending on whether or not the agent has won the respective Go-match. In our tunneling simulation, we designed a hierarchical point system that values (i.e. rewards/penalizes) either a state or an action. The rewarding/penalizing is done based on a list of conditions, that is worked through from top to bottom after every step of an episode and the agent is given the first reward where the condition is fulfilled. Except for the reward for achieving a breakthrough, all rewards are negative (i. Fig. 6. 3D visualization of the 2 × 2100 × 2 hypermatrix that represents the state of the simulation. G.H. Erharter et al.\n[page 7] Automation in Construction 127 (2021) 103701 7 e. penalties) and from top to bottom the penalties are sorted from biggest to smallest. It is therefore possible that multiple conditions are fulfilled in one state, but only the most severe penalty is given. Table 4 shows the list of rewards and their respective conditions. We designed the list of rewards based on practical engineering experience and requirements arising from training a completely unex­ perienced agent. For example, we penalize face instabilities higher than a too far distance between top heading and bench excavation, as the immediate safety hazard of an unstable face is bigger. In contrast to that, an even bigger penalty (i.e. -6 points) is triggered if the agent would try to do bench excavation ahead of the top heading excavation. A geo­ technician in real world tunneling would know from the start that doing so brings numerous technical-, logistical- and safety problems, but an untrained RL-agent does not. During the experiments for this study we tried to design the reward system in a way that it is the sum of all penalties that would be fulfilled in a given state and not a hierarchical system. However, doing this was not beneficial for training and seemed to confuse the agent as it apparently did not know what it was punished for. For example, a combination of the penalty for a too far distance between top heading and bench (i.e. -3) plus the penalty for using face support (i.e. -2) would sum up to −5 which is the same penalty as the one for unstable face conditions. As given in Section 2.1, the evaluation to check if a newly excavated face is stable or not, is done based on the face pressure equation for open face tunneling from [29] (Eq. (3)). In Eq. (3), pf is the required pressure to achieve stable face conditions, where a pf < 0 indicates stable con­ ditions and a pf ≥0 indicates unstable conditions, respectively the amount of pressure that is necessary to stabilize the face. γR is the ground’s unit weight, D the (equivalent) tunnel diameter (see chapter 3.1), d the advance length (describing the unsupported area of the un­ lined wall), c′ the effective cohesion and φ′ the effective friction angle. pf = γRD* ⎛ ⎜ ⎜ ⎜ ⎝ 2 + 3* ( d D )6*tanφ ′ 18*tanφ ′ −0.05 ⎞ ⎟ ⎟ ⎟ ⎠− c ′ tanφ ′ (3) 4. Training In RL, one episode is the whole succession of states in between an initial and a terminal state [26]. In other words, an episode is one whole match of a game, or in this simulation, one whole sequence of actions that ultimately should lead to a breakthrough of the tunnel. We started training with an ε of 1 (i.e. “initial exploration” in Table 2) to promote exploration in the initial phase of training. The exploration decay of 0.99997 (see Table 2) that decreases ε following Eq. (2) was determined by trial and error. A smaller exploration decay (i.e. faster reaching of the final exploration) has shown to increase in­ stabilities in the training process at an early stage which are presenting themselves in spontaneous increases of the loss and decreases of the reward. With the given exploration decay, the final exploration is reached after 99,858 episodes of training. After this point ε is kept at a constant value of 0.05. Training does not need to be aborted after reaching the final exploration. In the current simulation, we aimed at training the agents for 120,000 episodes to observe one full epsilon decay and some episodes beyond that to check for stable conditions in different rockmasses, as illustrated in Fig. 7 (except for cases where training became unstable at some point; see next section)."}], "source_title": "Reinforcement learning based process optimization and strategy development in conventional tunneling", "source_doi": "10.1016/j.autcon.2021.103701", "source_url": "https://www.sciencedirect.com/science/article/pii/S0926580521001527", "source_license": "cc-by", "zotero_item_key": "N5Q33JID"}
{"id": "26cf4becabb02256", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Spectral Angle Mapping and AI Methods Applied in Automatic Identification of Placer Deposit Magnetite Using Multispectral Camera Mounted on UAV; DOI: 10.3390/min12020268; license: cc-by-sa."}, {"role": "assistant", "content": "Another common occurrence is 3D laser scanning of mines in order to evaluate the current state of Minerals 2022, 12, 268. https://doi.org/10.3390/min12020268 https://www.mdpi.com/journal/minerals\n[page 2] Minerals 2022, 12, 268 2 of 19 a mine, which in turn enables better planning for the future [1–4]. In terms of rocks and minerals, the application of hyperspectral imaging in their classification based on their spectral characteristics has received positive feedback [5–9]. Even with its positive benefits, the employment of hyperspectral imaging has not been without imperfections. These include the time [9], monetary [8] and computational requirements necessary to process this large data [5]. Given these demands, it is difficult to perform rapid field spectral imaging as it takes time and sophisticated computational resources to acquire and process the hundreds of spectral bands. Part of this reason being, the need to process multitudes of redundant bands which do not necessarily improve the overall subject classification capabilities [7,10]. Coupled with the weight constraint, it is even more challenging to employ this technology in aerial investigations of a site at considerable spatial areas and resolutions [9]. On the other hand, research has shown that the application of a few specialized multispectral imaging bands in the classification rocks and minerals, offers similar, high subject resolving powers [11]. Specialized multispectral bands whose dimensionality has been reduced from hyperspectral imaging eliminate the said hyperspectral imaging disadvantages [12], for this reason, this study employs a 6-bands DJI Phantom 4 (P4) multispectral drone. Studies aimed at employing unmanned automated vehicles (UAVs) such as drones to identify certain rocks or minerals in mines are important as they offer a plethora of advantages [13,14]. Despite this, it is difficult to perform a thorough mine site rock or mineral assessment via a visible light camera mounted drone. This is due to the fact that such cameras do not collect the intrinsic characteristics of rocks and minerals beyond the visible light spectrum [6,15]. Mohajane et al. [4] applied a method which incorporated remote collection of vegeta- tion data in order to identify certain vegetation species via a multispectral camera mounted UAV drone. This study modifies and improves on this methodology, hence, a novel method by which rocks and/or minerals can be identified in rapid, large scale areas with detailed field assessments was born. This proposed method combines multispectral UAV drone technology with Spectral Angle Mapping (SAM), ML and DL algorithms to automatically identify magnetite iron sands. Potential advantage of this proposed system are; The UAV possesses multiple spectral sensors which all capture information pertaining to a scene at the same time; this results in reduced exposure times for each image, which in turn ensures one captures cleaner images without blur, and less heating of the sensors. Since the UAV drone spectral sensor comes pre-set with a low number of spectral dimensionalities to process, applying SAM becomes an advantage as it does not require the UAV to be specialized in magnetite identification. SAM takes advantage of the pre-existing spectral bands by assessing the difference between a reference subject’s (magnetite) spectral characteristics with those similar or different to it in an image scene [16,17]. AI algorithms such as ML and DL possess great spectral image classification power [18,19], allowing one to resolve qualitative as well as quantitative data in an objective and understandable manner, making it highly eligible in multiple industry applications. This study’s integrated system has the potential to enhance numerous aspects of the mining chain, hence the motivation for this study. The main objective of this study is to evaluate an effective approach for the automatic identification of magnetite iron sands. For this purpose, investigations on the feasibility of a multispectral camera mounted UAV drone will be discussed. Positive results of this study, assessed via SAM and AI models, will aid in ascertaining feasibility for field applications. The Study Area Kamaiso, the study area, is located along the west coast of Yamagata prefecture bounded to the west by the Japan Sea. This site is a magnetite iron sand placer deposit which is said to have resulted from volcanic activity by Mt."}], "source_title": "Spectral Angle Mapping and AI Methods Applied in Automatic Identification of Placer Deposit Magnetite Using Multispectral Camera Mounted on UAV", "source_doi": "10.3390/min12020268", "source_url": "https://www.mdpi.com/2075-163X/12/2/268/pdf?version=1645347690", "source_license": "cc-by-sa", "zotero_item_key": "VPIE7856"}
{"id": "704374fd43ea125c", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Reinforcement learning based process optimization and strategy development in conventional tunneling; DOI: 10.1016/j.autcon.2021.103701; license: cc-by."}, {"role": "assistant", "content": "11 it can be seen that within individual sequences of top heading or bench excavation, the agent alternates between supported and unsupported blasts which shows that it has successfully learned to avoid excessive use of support ahead of the face. 6. Conclusion and implications for digital tunneling Before drawing conclusions from the experiments, it should be pointed out that in this initial study, the agent’s possibilities to find creative and not yet imagined solutions to real world tunneling problems are confined to the given set of actions and the taken simplifications (see Section 2.1). We therefore see it as a success that the agent optimizes the given scenario and finds strategies that are comparable to current practices in tunneling. As given in the introduction, the current study should serve as a base for future developments of RL in tunneling. Consequently, optimizations that improve the current practices of “real life” tunneling are to be expected from future studies. In the experiments of the previous chapter, we can observe that the agents have found policies that minimize the overall amount of neces­ sary blasts, minimize changes between top heading and bench excava­ tion, favor long over short advance lengths and minimize the use of face support. These strategies show that the agents have learned to work in an efficient and economically optimized way. We see similarities in this RL-based tunneling strategy to real world tunneling paradigms like the NATM [49] which uses partial excavation to minimize the necessary support. Minimizing the number of changes between top heading or bench excavation while at the same time not exceeding a too long dis­ tance between them is also part of NATM tunneling as this optimizes construction site logistics on the one hand and safety requirements on the other. The found strategies that rely on support ahead of the face show similarities to the “Adeco” method [25] which uses heavy support installation and long advance lengths to deal with the encountered rockmass conditions. While adhering to safety requirements is impera­ tive, most “real life” tunneling methods work towards minimizing the number of necessary blasts which is a policy that was found by all the agents. TunnRL (see Section 2) has shown that it is not only a functioning environment/simulation of conventional tunneling, but also that a RL agent can successfully interact with it and learn optimized and inno­ vative strategies that seem realistic compared to real world tunnel excavation. At the same time, we see the challenges of computational instabilities and trial and error approach in the process of developing well-functioning models, highlighting this early stage in RL for tunneling. Clearly both the reward system, the rockmass-environment and the action-system has room for improvement. Still we see a signif­ icant potential in the TunnRL-concept: firstly, for an on-face decision support system in a further developed and more realistic version, and secondly as a first step to more advanced automation in underground Table 5 Statistics of the training runs presented in Fig. 7. Values were computed from the 500-episode sliding window average and the episode of the respective value is given in parenthesis behind it. Row-wise best performances are highlighted in grey. G.H. Erharter et al.\n[page 10] Automation in Construction 127 (2021) 103701 10 construction. Where many of the developments of tunnel processes today address the automation of small-scale processes in the tunnel or at the excavation face, a further developed TunnRL could be part of the main controlling mechanism that operates the overall tunnel construc­ tion site. Albeit the agents in this simulation were not able to find new and undiscovered strategies of tunnel excavation (see previous chapter), we see the fact that completely untrained agents are able to find tunneling strategies which are comparable to “real world tunneling” as a proof of concept that RL is successfully applicable to this kind of prob­ lem. The main goal of the study is therefore fulfilled, and future studies will work towards giving the agent more capabilities and increasing the environment’s realism. This will ultimately pave the way for an opti­ mized decision finding process in sequential tunneling. 7. Outlook Albeit we designed TunnRL in a practice related context, there are numerous improvements to make the framework more realistic and more robust."}], "source_title": "Reinforcement learning based process optimization and strategy development in conventional tunneling", "source_doi": "10.1016/j.autcon.2021.103701", "source_url": "https://www.sciencedirect.com/science/article/pii/S0926580521001527", "source_license": "cc-by", "zotero_item_key": "N5Q33JID"}
{"id": "33c00e8034ea0698", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Detection of Ionospheric Scintillation Based on XGBoost Model Improved by SMOTE-ENN Technique; DOI: 10.3390/rs13132577; license: cc-by-sa."}, {"role": "assistant", "content": "The training set is sampled to obtain multiple subsets, and then, several decision trees with the same structure are constructed, each of which is trained separately by using the constructed training subsets. The classification result depends on votes from all the decision trees trained on the testing set. More details can be found in the literature [16]. Consisting of different decision trees, the core of the XGBoost algorithm is learning a new function f (x) to fit the last predicted residuals, adding decision trees continually and continuing to split features to grow a tree before meeting the growth conditions. After completing the training process and obtaining k trees, we needed to predict the novel data. According to the characteristics of samples, they will be classified to a leaf node corresponding to a score in each tree, and the sum of the scores for each tree to be regarded as the predicted value of the sample. Obviously, the goal of the algorithm is to make the\n[page 8] Remote Sens. 2021, 13, 2577 8 of 22 predicted value of all the trees ˆyi as close as possible to the corresponding true value yi, with as much generalization capability as possible. Based on the current tree, an additional tree is added to fit the residual between the predicted result of the previous trees and corresponding true value. To choose the next tree to be added, we introduced the objective function combined with the loss function and regular function, as shown in (6). Obj(t) = n ∑ i=1 l \u0010 yi, ˆyi(t−1) + ft(xi) \u0011 + Ω( ft) + cons (6) where i represents the ith sample, n is the number of samples, yi is the true score of the current tree, and ˆyi(t−1) refers to the predicted value of t −1 trees. ft is a new function of the current tree; Ω( ft) is the corresponding regularization item. Additionally, the constant value cons can be ignored without affecting the following objective process. The objective function can be approximated as follows after using Taylor expansion: Obj(t) = n ∑ i=1 \u0014 l \u0010 yi, ˆyi(t−1)\u0011 + gi ft(xi) + 1 2hi ft2(xi) \u0015 + Ω( ft) + cons = n ∑ i=1 \u0014 giωq(xi) + 1 2hiωq2(xi) \u0015 + γT + λ1 2 T ∑ j=1 hiωj 2 + C (7) where gi = ∂ˆy(t−1)l \u0010 yi, ˆy(t−1)\u0011 ; hi = ∂2 ˆy(t−1)l \u0010 yi, ˆy(t−1)\u0011 , T is the number of the leaf nodes, and ωj is the weight of the jth leaf node. γ and λ are used to control the complexity of the trees. Defining Gj = ∑ i∈Ij gi and Hj = ∑ i∈Ij hi, the optimal solution ω∗ j = − Gj Hj+λ; then, the optimal objective function without a constant value C can be obtained as follows. Obj(t) = T ∑ j=1 \u0014 Gjωj + 1 2 Hjωj 2 \u0015 + γT = −1 2 T ∑ j=1 Gj Hj + λ + γT (8) Furthermore, the greedy algorithm is utilized to enumerate all the different tree structures to find the optimal splitting node, achieving the maximum gain of the objective function after splitting, as shown in (9). Gain = 1 2 \" G2 L HL + λ + G2 L HL + λ −(GL + GR)2 HL + HR + λ # −γ (9) where GL and GR are the first-order gradient values of the leaf nodes on left and right after splitting. Similarly, HL and HR are corresponding second gradient values. A brief schematic diagram of the XGBoost algorithm is shown in Figure 3. There are some preset parameters used to control the tree building, such as γ, λ, the maximum depth of the tree, the minimum weight of the child node, and the learning rate for each tree. The main purpose is to limit the complexity and weight of each tree, so that the overfitting problem can be mitigated or prevented. 3.2. SMOTE-ENN Resampling Technique An imbalance in the dataset refers to the phenomenon of certain types of samples being lower in number than other types, and researchers always have more interest in the minority class samples [42]. For similar characteristics of data, many classification algorithms have difficulty in accurately expressing the inherent features because of the lack of information on the minority class samples, which causes the decision boundary to be greatly compressed in the classification system. Although the overall accuracy of the detection model is excellent, it is unable to effectively detect the minority target samples that ought to be detected. As shown in Table 2, the weak, moderate and strong scintillation\n[page 9] Remote Sens. 2021, 13, 2577 9 of 22 events are marked as 0, 1 and 2, respectively. When the training sample proportion ranges from 1:1:1 to 7:7:1 based on a constant 100,000 entries, the overall accuracy remains relatively stable and good, ranging from 90.18% to 93.32%."}], "source_title": "Detection of Ionospheric Scintillation Based on XGBoost Model Improved by SMOTE-ENN Technique", "source_doi": "10.3390/rs13132577", "source_url": "https://www.mdpi.com/2072-4292/13/13/2577/pdf?version=1625212202", "source_license": "cc-by-sa", "zotero_item_key": "4YEIWMK3"}
{"id": "469b1fbfa586c9f7", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Spectral Angle Mapping and AI Methods Applied in Automatic Identification of Placer Deposit Magnetite Using Multispectral Camera Mounted on UAV; DOI: 10.3390/min12020268; license: cc-by-sa."}, {"role": "assistant", "content": "Chokai, an event estimated to have taken place 600,000 years ago [20]. It makes sense to explore the possibility of mining\n[page 3] Minerals 2022, 12, 268 3 of 19 easily accessible placer deposits such as those in Kamaiso, especially with the worldwide decline in ore grades [21]. These potential resources present an opportunity to maximize profit margins depending on the amounts present. Deposits located at coastal areas are rather difficult to quantify due to their close proximity with the sea. Air borne methods such as the employment of UAV drones present an unconventional way of exploring and quantifying target resources within this area effectively [22]. A map showing a 30 m × 90 m area from which experiments were conducted in an attempt to apply UAV drone technology in identifying magnetite iron sands is shown in Figure 1. Figure 1. Study area map showing Kamaiso magnetite iron sand placer deposits. Red polygon denotes the 30 m by 90 m study area within which the UAV drone shown was flown. Map acquired from: Google maps (2021). 2. Methodological Strategies for SAM and AI Integration 2.1. Combination of UAV Drone Technology with Multispectral Imaging Air borne drone technology has evolved over the years, with recent developments catering for specialized industrial applications. The protection of human life, as sites can be assessed from a remote area, is one of the major advantages of UAV drone technology [3,22]. In addition, camera-mounted drones aid in real time assessments of sites via visual feed. This footage can be in the form of real, false or pre-processed colours which are meant to represent a certain occurrence understood by the operator [4]. It is supposed that hyperspectral imaging is perhaps one of the best methods by which rocks and minerals can be inferred through their spectral characteristics [5]. However, van de Meer et al. [7] have been able to achieve high classification accuracy outputs in similar rock identification problems via multispectral imaging. Motivated by the demonstrated multispectral capabilities this paper employs a DJI P4 Multispectral camera mounted UAV drone (manufactured by Da-Jiang Innovations, Shenzhen, China) released in September 2019 (Figure 2). To achieve centimeter-level position accuracy, the drone has a real time kinematics (RTK) module attached to it. In addition, the six cameras are 1/2.9-inch (com- plementary metal oxide semiconductor) CMOS sensors, all at 2 megapixels with a global shutter, on a 3-axis stabilized gimbal to ensure accurate multispectral image capturing at spatial resolutions as high as 0.1 cm/pixel. The spectral sensor specifications are as follows: Blue: 450 nm ± 16 nm, Green: 560 nm ± 16 nm, Red: 650 nm ± 16 nm, Red Edge:\n[page 4] Minerals 2022, 12, 268 4 of 19 730 nm ± 16 nm, Near Infrared: 840 nm ± 26 nm, which are all captured as TIFF image files. The drone moreover boasts a 62.7◦field of view, a net weight of 468 g and an RBG camera which facilitates the extraction of magnetite iron sand spectral information in the visible light spectral range of 400–700 nm as a JPG image. These attributes subsequently make the system effective in a wide range of electromagnetic spectra within the visible-near-infrared range (VNIR) with a total of 6 multispectral bands. Figure 2. The 6-bands multispectral sensor with a 62.7◦field of view; (1) Blue: 450 nm ± 16 nm, (2) Green: 560 nm ± 16 nm, (3) Red: 650 nm ± 16 nm, (4) Red Edge: 730 nm ± 16 nm, (5) Near Infrared: 840 nm ± 26 nm and (6) RGB camera: 400–700 nm represent the multispectral sensor capabilities. Images captured are corrected for parallax error via an edge detection method. The red point is the reference area from which magnetite iron sands are known to exist. The UAV drone flight elevation plan was set to 3 different variations so as to explore the effect of ground resolution, proximity to the subject, and the ease of flying. From examining these attributes, it would be easier to find the optimal elevation from where magnetite iron sand identification can be performed effectively. The 3 UAV flight elevations measured from the camera sensor to the subject were 2 m, 10 m and 20 m. The drone was moreover programmed such that the aerial multispectral imaging procedure is automat- ically performed at intervals that ensured an equal amount of area (30 m × 90 m) was captured by the end of each flight mission."}], "source_title": "Spectral Angle Mapping and AI Methods Applied in Automatic Identification of Placer Deposit Magnetite Using Multispectral Camera Mounted on UAV", "source_doi": "10.3390/min12020268", "source_url": "https://www.mdpi.com/2075-163X/12/2/268/pdf?version=1645347690", "source_license": "cc-by-sa", "zotero_item_key": "VPIE7856"}
{"id": "2420401c974ea5db", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Intelligent Classification of Surrounding Rock of Tunnel Based on 10 Machine Learning Algorithms; DOI: 10.3390/app12052656; license: cc-by-sa."}, {"role": "assistant", "content": "Drilling parameter box diagrams for each surrounding rock grade. (a) Penetration velocity (Vp, m/min); (b) hammer pressure (Ph, bar); (c) feed pressure (Pf, bar); (d) rotation pressure (Pr, bar). According to the box diagram method, when the value is not within the interval represented by the following formula Equation (1), it is regarded as an outlier [Q1 −1.5 × IQR, Q3 + 1.5 × IQR] (1) where Q1 is the first quartile, namely the equal to the 25th percentile of all values in the sample from small to large, Q3 is the third quartile, namely the equal to the 75th percentile of all values in the sample from small to large, IQR is the interquartile range, namely IQR is equal to Q3 minus Q1. In Figure 6, under different surrounding rock grades, all four drilling parameters contained outliers. To eliminate the effects of the outliers, the average value of each parameter was used to replace the outlier of each surrounding rock in this study. The changes in data characteristics before and after cleaning are compared in Table 4.\n[page 9] Appl. Sci. 2022, 12, 2656 9 of 20 Table 4. Comparison of the data characteristics before and after cleaning. Surrounding Rock Grade Index Vp Ph Pf Pr Before After Before After Before After Before After III mean 2.35 2.04 144.32 148.85 66.95 67.69 91.91 92.18 std 0.97 0.51 12.59 5.98 8.02 5.91 10.29 9.69 min 1.08 1.08 99.95 132.88 29.05 53.70 46.94 67.27 max 5.56 3.39 165.41 165.41 91.45 81.44 112.97 112.97 IV mean 3.19 3.19 135.06 135.06 63.17 62.86 79.68 79.25 std 1.03 1.03 14.00 14.00 13.21 12.67 11.24 9.59 min 0.66 0.66 89.79 89.79 35.97 35.97 48.84 55.38 max 5.49 5.49 161.38 161.38 107.41 97.95 113.53 102.29 V mean 3.47 3.54 124.96 125.87 51.75 50.72 73.18 72.62 std 0.85 0.67 11.68 9.85 11.06 9.59 7.59 5.65 min 0.83 1.97 90.82 101.39 27.43 27.43 53.47 57.70 max 6.35 5.05 155.88 150.65 83.75 73.60 101.17 86.72 Note: The full name of each index and corresponding abbreviated in the table is the mean value (mean), the standard deviations (std), the minimum value (min) and the maximum value (max). In Table 4, after cleaning, under different surrounding rock grades, the distributions of these drilling parameters are more centralized. Specifically, the maximum value is smaller, the minimum value is larger, and the standard deviation is smaller. 2.4. Sample Imbalance Treatment Referring to machine learning classification, when the difference in the number of samples of all the classes is too large, the training model will pay too much attention to the sample characteristics with a greater proportion, so the classification effect of the samples with a smaller proportion is not ideal. This is called the problem of sample imbalance. These data sets cover three classes of surrounding rock (namely, grade-III, -IV, and -V), and each class contains 325, 420, and 167 samples belonging to the imbalance sample set. There are three common approaches to deal with unbalanced samples in machine learning: 1. Over-sampling The over-sampling method achieves sample balance by increasing the number of minority samples in the classification. The most direct method is to simply copy the minority samples to form multiple records. The improved over-sampling methods produce new composite samples by adding random noise, interference data to a few classes, or certain rules such as the synthetic minority over-sampling technique (SMOTE) and adaptive synthetic sampling (ADASYN). 2. Under-sampling The under-sampling method achieves sample balance by reducing the number of majority samples in the classification. The most direct method is to randomly remove some majority class samples. The disadvantage of this method is that some important information from the majority class samples may be lost. 3. Sample weight adjustment This method is used to guide models to learn more features of the minority samples in the classification by assigning higher weights to them during the machine learning model training. Generally, the weights of various samples are set to be inversely proportional to the sample size. Considering the small size of the sample data set, to sufficiently learn the characteristics of all types of samples during the machine learning model training and to improve the universality of the model, the SMOTE over-samples to replenish samples in grade-III and grade-V of the surrounding rock. The scatter diagram of the data set after over-sampling is shown in Figure 7.\n[page 10] Appl. Sci."}], "source_title": "Intelligent Classification of Surrounding Rock of Tunnel Based on 10 Machine Learning Algorithms", "source_doi": "10.3390/app12052656", "source_url": "https://www.mdpi.com/2076-3417/12/5/2656/pdf?version=1646383280", "source_license": "cc-by-sa", "zotero_item_key": "UGAERXYU"}
{"id": "b796f260fc129c17", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Intelligent Classification of Surrounding Rock of Tunnel Based on 10 Machine Learning Algorithms; DOI: 10.3390/app12052656; license: cc-by-sa."}, {"role": "assistant", "content": "Back-Analysis Method for Stope Displacements Using Gradient-Boosted Regression Tree and Firefly Algorithm. J. Comput. Civil. Eng. 2018, 32, 04018031. [CrossRef] 14. Zhang, L.; Wu, X.; Ji, W.; Abourizk, S.M. Intelligent Approach to Estimation of Tunnel-Induced Ground Settlement Using Wavelet Packet and Support Vector Machines. J. Comput. Civil. Eng. 2016, 31, 04016053. [CrossRef] 15. Zhou, C.; Ding, L.; Zhou, Y.; Zhang, H.; Skibniewski, M.J. Hybrid Support Vector Machine Optimization Model for Prediction of Energy Consumption of Cutter Head Drives in Shield Tunneling. J. Comput. Civil. Eng. 2019, 33, 04019019. [CrossRef] 16. Xue, Y.; Bai, C.; Qiu, D.; Kong, F.; Li, Z. Predicting rockburst with database using particle swarm optimization and extreme learning machine. Tunn. Undergr. Space Technol. 2020, 98, 103287. [CrossRef] 17. Zhao, H.; Ru, Z.; Chang, X.; Yin, S.; Li, S. Reliability analysis of tunnel using least square support vector machine. Tunn. Undergr. Space Technol. 2014, 41, 14–23. [CrossRef] 18. Liu, W.; Cai, L.; Chen, J.; Wang, Y.; Wu, H. Reliability Analysis of Operational Metro Tunnel Based on a Dynamic Bayesian Copula Model. J. Comput. Civil. Eng. 2020, 34, 05020002. [CrossRef] 19. Leu, S.-S.; Chen, C.-N.; Chang, S.-L. Data mining for tunnel support stability: Neural network approach. Automat. Constr. 2001, 10, 429–441. [CrossRef] 20. Liu, K.; Liu, B. Optimization of smooth blasting parameters for mountain tunnel construction with specified control indices based on a GA and ISVR coupling algorithm. Tunn. Undergr. Space Technol. 2017, 70, 363–374. [CrossRef] 21. Liu, J.; Jiang, Y.; Ishizu, S.; Sakaguchi, O. Estimation of tunnel support pattern selection using artificial neural network. Arab. J. Geosci. 2020, 13, 1–12. [CrossRef] 22. Rana, A.; Bhagat, N.; Jadaun, G.; Rukhaiyar, S.; Pain, A.; Singh, P. Predicting Blast-Induced Ground Vibrations in Some Indian Tunnels: A Comparison of Decision Tree, Artificial Neural Network and Multivariate Regression Methods. Mining. Metall. Explor. 2020, 37, 1039–1053. [CrossRef] 23. Yazdani-Chamzini, A. Proposing a new methodology based on fuzzy logic for tunnelling risk assessment. J. Civ. Eng. Manag. 2014, 20, 82–94. [CrossRef] 24. Liu, Y.; Chen, H.; Zhang, L.; Wang, X. Risk prediction and diagnosis of water seepage in operational shield tunnels based on random forest. J. Civ. Eng. Manag. 2021, 27, 539–552. [CrossRef] 25. Chen, Z.; Zhang, Y.; Li, J.; Li, X.; Jing, L. Diagnosing tunnel collapse sections based on TBM tunneling big data and deep learning: A case study on the Yinsong Project, China. Tunn. Undergr. Space Technol. 2020, 108, 103700. [CrossRef] 26. Wen, Z.; Rong, X.; Wang, Z.; Han, S.; Shi, Y. A Fast Estimation Method of Soil Discharged by an Earth Pressure Balanced Shield Machine. KSCE J. Civ. Eng. 2021, 25, 2239–2249. [CrossRef] 27. Hu, M.M.; Li, B.; Zhang, B.; Wang, R.R.; Chen, L. Improved SVR Method for Predicting the Cutting Force of a TBM Cutter Using Linear Cutting Machine Test Data. KSCE J. Civ. Eng. 2021, 25, 4425–4442. [CrossRef]\n[page 20] Appl. Sci. 2022, 12, 2656 20 of 20 28. Jung, J.H.; Chung, H.; Kwon, Y.S.; Lee, I.M. An ANN to Predict Ground Condition ahead of Tunnel Face using TBM Operational Data. KSCE J. Civ. Eng. 2019, 23, 3200–3206. [CrossRef] 29. Lee, S.S. Prediction of Subsidence during TBM Operation in Mixed-Face Ground Conditions from Realtime Monitoring Data. Appl. Sci. 2021, 11, 12130. [CrossRef] 30. Xu, H.; Zhou, J.; Asteris, P.G.; Armaghani, D.J.; Tahir, M.M. Supervised Machine Learning Techniques to the Prediction of Tunnel Boring Machine Penetration Rate. Appl. Sci. 2019, 9, 3715. [CrossRef] 31. Zafirovski, Z.; Jovanovski, M.; Moslavac, D.; Krakutovski, Z. Application of Neural Networks in Analyzing of Rock Mass Parameters in Tunneling; ResearchGate: Split, Croatia, 2014. 32. You, K. A case study on the utilization of tunnel face mapping data for a back analysis based on artificial neural network. KSCE J. Civ. Eng. 2014, 18, 751–759. [CrossRef] 33. Jalalifar, H.; Mojedifar, S.; Sahebi, A.; Nezamabadi-Pour, H. Application of the adaptive neuro-fuzzy inference system for prediction of a rock engineering classification system. Comput. Geosci. 2011, 38, 783–790. [CrossRef] 34. Hasegawa, N.; Hasegawa, S.; Kitaoka, T.; Ohtsu, H. Applicability of Neural Network in Rock Classification of Mountain Tunnel. Mater. Trans. 2019, 60, 758–764. [CrossRef] 35. Lee, H.L.; Kim, J.S.; Hong, C.H.; Cho, D.K."}], "source_title": "Intelligent Classification of Surrounding Rock of Tunnel Based on 10 Machine Learning Algorithms", "source_doi": "10.3390/app12052656", "source_url": "https://www.mdpi.com/2076-3417/12/5/2656/pdf?version=1646383280", "source_license": "cc-by-sa", "zotero_item_key": "UGAERXYU"}
{"id": "2c8fdcd1ea8f578c", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Detection of Ionospheric Scintillation Based on XGBoost Model Improved by SMOTE-ENN Technique; DOI: 10.3390/rs13132577; license: cc-by-sa."}, {"role": "assistant", "content": "Thus, based on the above feature set and non-Gaussian noise characteristics of ionospheric scintillation [40], the other four metrics of variances and covariance are also presented to further reflect the features. σSI = 1 N −1 N ∑ n−1 (SIn −⟨SI⟩)2 (1) cov \u0010 I2, Q2\u0011 = 1 N −1 N ∑ n−1 \u0010 In2 − D I2E\u0011\u0010 Qn2 − D Q2E\u0011 (2) σI = 1 N −1 N ∑ n−1 (In −⟨I⟩)2 (3) σQ = 1 N −1 N ∑ n−1 (Qn −⟨Q⟩)2 (4) where σSI, σI and σQ are variance of SI, I and Q, respectively, and cov \u0000I2, Q2\u0001 is the covariance of I2 and Q2. N is the number of samples used in one value. To remove the impact of thermal noise and reserve the characteristics of the scintillation phenomenon, the frequency of the initial observables I and Q was f = 1000 Hz. For each feature, the observation window was set as T = 60 s, meaning that there were N = T· f = 60000 initial observables averaged in the observation period. The frequency of observation was 1 Hz, which was in line with the shifting window of 1 s. Based on that, the features used in this work are described as follows. X = n ⟨I⟩, ⟨Q⟩, D I2E , D Q2E , ⟨SI⟩, D SI2E , σSI, cov \u0010 I2, Q2\u0011 , σI, σQ o (5)\n[page 7] Remote Sens. 2021, 13, 2577 7 of 22 The scintillation phenomenon may persist for several hours or longer, meaning that the data will be affected by the scintillation for a significant period of time rather than several minutes or less. The scintillation events were manually marked as L = {0, 1, 2} based on the visual inspection of thresholds, divided into 0 < S4 ≤0.2, 0.2 < S4 ≤0.6 and 0.6 ≤S4, representing weak, moderate and strong scintillation events marked as 0, 1 and 2, respectively. Meanwhile, the corresponding carrier-to-noise density power ratio (C/N0) was also taken into consideration in the manual detection process. This manual labelled approach combined with personal knowledge and experience can reserve the transient phases of the events and reduce the missed detections, significantly enhancing the labelling accuracy and detection performance. 3. Methodology The goal of this approach was to propose an improved eXtreme Gradient Boosting (XGBoost) algorithm combined with the synthetic minority oversampling technique and edited nearest neighbor (SMOTE-ENN), comparing the performance with that of the decision tree and random forest algorithms, which have been successfully used in recent research [16]. 3.1. XGBoost Algorithm The XGBoost algorithm was proposed by Chen Tianqi in 2016, presenting low com- putational complexity, a fast running speed and high accuracy [41]. As it is an inefficient ensemble learning algorithm, the boosting is aimed at transforming a weak classifier into a strong classifier to achieve good accuracy. Moreover, the gradient boosting attempts to improve robustness by making the algorithm’s loss function drop along its gradient direction in the iteration process. Additionally, as a fast implementation of the gradient boosting algorithm, XGBoost can make full use of multi-core CPUs for parallel computation and improve the accuracy, significantly reducing the computational loads and enhancing the accuracy compared with other widely used algorithms such as the decision tree and random forest. As the basis function of XGBoost, decision tree-based solutions for classification tasks have been used successfully in various fields, and corresponding advanced algorithms based on that are gradually being proposed to enhance the classification performance in terms of accuracy, precision, generalization and computing efficiency. There are three typically used decision-tree algorithms based on information theory: ID3, C4.5 and classifi- cation and regression tree (CART); the classification regulation of CART is derived from the Gini index, and CART is used more frequently than the two other algorithms. It not only assigns categories to leaf nodes, but also considers the possibility of all attributes being selected as leaf nodes. When all the samples in the node belong to the same attribution or the depth of the decision tree reaches a preset threshold, the tree construction stops. Due to the overfitting phenomenon for the former condition, it is vital to set the max depth value before training. Moreover, the random forest consists of multiple structurally similar decision trees for determining the dataset together to prevent overfitting and reduce the variance of an estimate."}], "source_title": "Detection of Ionospheric Scintillation Based on XGBoost Model Improved by SMOTE-ENN Technique", "source_doi": "10.3390/rs13132577", "source_url": "https://www.mdpi.com/2072-4292/13/13/2577/pdf?version=1625212202", "source_license": "cc-by-sa", "zotero_item_key": "4YEIWMK3"}
{"id": "e13616940d44c8bf", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Spectral Angle Mapping and AI Methods Applied in Automatic Identification of Placer Deposit Magnetite Using Multispectral Camera Mounted on UAV; DOI: 10.3390/min12020268; license: cc-by-sa."}, {"role": "assistant", "content": "These outputs, except time taken to train the models, moreover govern industrial applicability, and the overall viability of a model in solving tasks at hand [6]. A compilation of the top three performing ML models for the three multispectral UAV drone flight elevations post SAM and ground truthing are shown in Table 2. The differences in classification capabilities based on flight elevations are demonstrated quantitatively. Global accuracy is defined as the validation accuracy acquired post model training. Average per-class precision refers to the discrete magnetite iron sands (target) and the non-magnetite sands (noise) classification score averages in testing the models. Lastly, training time refers to time taken to train the models to classify the datasets based on the existence or absence of magnetite iron sand. These performance variables are important in comparing the different ML as well as DL models given their intended application.\n[page 13] Minerals 2022, 12, 268 13 of 19 Table 2. Top 3 machine learning classification model comparisons at 2 m, 10 m and 20 m UAV flight elevations based on 0.12, 0.13 and 0.17 SAM cosine similarity thresholds, respectively. Reprinted with permission from ref. [1]. 2021 AusIMM [1] (p. 40). UAV Drone Flight Elevation Machine Learning Model Global Accuracy (%) Average Per-Class Precision (%) Training Time (Seconds) 2 m Ensemble (Bagged Trees) 78.6 83.4 8.4 Ensemble (Subspace KNN) 71.4 77.8 8.1 Ensemble (RUS Boosted Trees) 85.7 84.5 5.8 10 m Tree (Fine-tree) 78.6 83.4 1.5 Tree (Medium-tree) 78.6 83.4 1.0 Tree (Course-tree) 78.6 83.4 0.9 20 m Tree (Fine-tree) 85.7 88.9 1.9 Tree (Medium-tree) 85.7 88.9 1.2 Tree (Course-tree) 85.7 88.9 1.0 At 2 m UAV flight elevation, Table 2 results dictate that the best performing model was Ensemble (RUS Boosted Trees). This model acquired a global accuracy of 85.7%, per-class precision of 84.5%, and a training time of 5.8 s. At 10 m UAV flight elevation, the Tree (Course-tree) model performed the best. This model acquired a global accuracy of 78.7%, per-class precision of 83.4%, and a training time of 1 s. Lastly, 20 m UAV flight elevation, the Tree (Course-tree) model performed the best. This model acquired the best outputs in terms of a global accuracy of 85.7%, per-class precision of 88.9%, and a training time of 1 s. The differences in accuracies acquired and elapsed times when training the ML models for the three UAV flight elevations can, moreover, be appreciated based on the results compiled in Table 2. The reason being, this confirms the initial hypothesis which stated that with the proposed automatic magnetite iron sand identification system, rapid run times and good accuracies are potentially attainable. This is, of course, without compromise to the intrinsic differences in the multispectral characteristics of magnetite iron sands and the surrounding non-magnetite beach sand. Figure 9 presents two performance metrics used to further assess the viability of the best performing ML models for the three UAV flight elevations. The first is True Positive Rates (TPR), this refers to the probability that an actual positive will test positive (Equa- tion (2)). The second is False Negative Rates (FNR), which are defined as the probability that a true positive will be misclassified by the test (Equation (3)). These two variables are highly viable in assessing the capability of the ML as well as DL models in classifying mag- netite iron sands. At 2 m, the ML model has an 80% TPR, which may not seem satisfactory considering that target data should ideally have a 100% target return. This means that 20% of the target pixels were misclassified as noise, suggesting that the model at 2 m did not perform well at discerning magnetite iron sand pixels from the rest of the non-target pixels. In essence, this means that operating and collecting data at this elevation may not be the best practice. At 10 m, the target has a 100% TPR, signifying that the target pixels were correctly classified; this moreover suggests that the target output did not include any noise. At 20 m UAV flight elevation, the same assessment can be made, suggesting that these flight elevations are the most viable out of the three. At 20 m however, the model has a higher noise TRP of 77.8% compared to the 67.7% at 10 m."}], "source_title": "Spectral Angle Mapping and AI Methods Applied in Automatic Identification of Placer Deposit Magnetite Using Multispectral Camera Mounted on UAV", "source_doi": "10.3390/min12020268", "source_url": "https://www.mdpi.com/2075-163X/12/2/268/pdf?version=1645347690", "source_license": "cc-by-sa", "zotero_item_key": "VPIE7856"}
{"id": "96a82794aa6e103f", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: How to evaluate performance of prediction methods? Measures and their interpretation in variation effect analysis; DOI: 10.1186/1471-2164-13-s4-s2; license: cc-by."}, {"role": "assistant", "content": "Method testing schemes Three approaches can be used for testing method per- formance and can be classified according to increasing reliability (Fig. 1). Challenges aim to test whether certain problems can be addressed with existing tools and to find out what kind of methods will be needed in the future. Critical Assessment of Structure Predictions (CASP) [6] was the first chal- lenge of this kind in biosciences. The idea was, and still is, even when CASP has been running for 20 years, to test how prediction methods behave on different protein structure related tasks. The method developers apply their systems without knowing the correct result (blind test), which however is available for the challenge asses- sors. This setup allows independent testing of method performance. In a similar vein, other critical assessment challenges have been organized e.g. for Critical Assess- ment of protein Function Annotation (CAFA) [7] and Critical Assessment of PRediction of Interactions (CAPRI) [8]. CAGI, Critical Assessment of Genome Interpretation (http://genomeinterpretation.org/), is a challenge for method developers in the field of phenotypic impacts of genomic variation. The second CAGI prediction season was organized during fall 2011. These challenges do not aim for systematic analysis of predictions, instead they assess what is currently doable, providing proof of con- cept, charting where to direct future efforts, and identi- fying new areas where predictive approaches would be needed. The second test strategy is typically used by method developers to test their approaches. These are usually done with developer collected test sets (especially when benchmark datasets are lacking) and report certain per- formance parameters. Most often the testing is not com- prehensive, and the results are incomparable with those obtained from other methods e.g. due to using different test sets. Sometimes evaluation parameters are selectively presented which leads to problems in determining the true merits and pitfalls of methods. The third approach, systematic analysis, uses approved and widely accepted benchmark dataset(s) and suitable evaluation measures to explain method performance. It is hoped that in the future the variation effect program developers would use benchmark test sets and compar- able measures. This is already the general practice e.g. in the multiple sequence alignment (MSA) field. Prediction methods for classification A plethora of pattern recognition methods have been applied to problems in bioinformatics including rule based, statistical methods and machine learning -based methodologies. The goal of machine learning is to train a computer system to distinguish i.e. classify cases based on known examples. Machine learning methods include sev- eral widely differing approaches such as support vector machines, neural networks, Bayesian classifiers, random forests and decision trees. In the following discussion we concentrate on machine learning methods as they are nowadays widely used to tackle complex phenomena, which would be otherwise difficult to handle. Successful machine learning method development requires good quality training set. The data- set should represent the space of possible cases. This space is huge for genetic variations as they can have so many different effects and underlying mechanisms. Another aspect is the choice of the machine learning approach. There is not a superior architecture among them. Third, the quality of the predictor depends on how the training has been done, which features are used to Figure 1 Method performance analysis schemes The performance of the computational methods can be addressed with three different approaches which yield different reliability for the assessment. Vihinen BMC Genomics 2012, 13(Suppl 4):S2 http://www.biomedcentral.com/1471-2164/13/S4/S2 Page 2 of 10\n[page 3] explain the phenomenon and optimization of the method. Fig. 2 depicts the principle underlying machine learning in a two-class classification task. The predictor is trained with known positive and negative instances in an approach called supervised learning. This leads to reorganization of the system, details of which differ according to the architec- ture employed. Once the method has learned to distinguish between the cases it can be applied to predict the class of unknown cases."}], "source_title": "How to evaluate performance of prediction methods? Measures and their interpretation in variation effect analysis", "source_doi": "10.1186/1471-2164-13-s4-s2", "source_url": "https://bmcgenomics.biomedcentral.com/counter/pdf/10.1186/1471-2164-13-S4-S2", "source_license": "cc-by", "zotero_item_key": "LKJJ9WPZ"}
{"id": "a25eabb43b7fa76d", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Reinforcement learning based process optimization and strategy development in conventional tunneling; DOI: 10.1016/j.autcon.2021.103701; license: cc-by."}, {"role": "assistant", "content": "Ground type Specific weight [kN/m3] Cohesion [kPa] Friction angle [◦] Permeability [m/s] gt1 24 23 20 10−5 gt2 25 40 30 10−5 G.H. Erharter et al.\n[page 5] Automation in Construction 127 (2021) 103701 5 The output of the third convolutional layer is flattened/vectorized before it is fed into the dense layer and we used the mean squared error as a loss function. 3.2.1. Actions The agent can choose from 8 different actions (Table 3). Possible actions are either top heading – or bench excavation with advance lengths of either 2 or 4 m. In each step the simulation carries out the following operations: ▪ the agent chooses and executes one of the 8 actions of Table 3 and the respective length of the generated geological section is revealed ▪ it is calculated/checked if the new position of the excavation face is within a stable area or not (details in Section 3.3.2) As given in Section 2.1, the support ahead of the face has the effect on the simulation that, there cannot be unstable conditions as long as the excavation face is within the supported area. Because of technical rea­ sons - mostly related to the reward system (see Table 4) - each action is assigned an “action code” (a). An a < 200 denotes top heading exca­ vation and a ≥200 bench excavation. The choice of actions of Table 3 is based on practical engineering experience on the one hand and technical limitations of the RL agent on the other hand. On the practical side, especially the advance length is Fig. 4. Top row: an exemplary unique geological section, where brown indicates weak (gt1) and blue stronger rock (gt2). The positions of the top heading and bench are at 165.0 m and 125.0 m respectively. Bottom row: the random walk that is used to generate the geological section. Values above 0.5 are converted to gt2 and below to gt1. Note that the x-axis is the tunnel length in decimeters which corresponds to the number of datapoints of the random walk. Table 2 Used hyperparameters for the DQN agent. Except for the exploration decay (see Section 3.2.1) all are identical in their meaning as the extended data Table 1 in [14]. Hyperparameter Value Replay memory size 100,000 Replay start size 1000 Minibatch size 64 Discount (γ) 0.99 Target network update frequency 10 Initial exploration (ε) 1 Final exploration (ε) 0.05 Learning rate 0.00025 Gradient momentum 0.95 Exploration decay (εd) 0.99997 Fig. 5. Schematic representation of the DQN agent’s ANN architecture. Note the visualization of rockmass-matrix and the support-matrix to the left. The numbers below each layer are the respective shape of the layer’s weights. Dashed connection lines between layers are only for illustrational purposes. Symbols at the output layer represent the eight possible actions (ordered as in Table 3) that are chosen via Q-values by the agent. Table 3 The eight possible actions the agent can choose from. Action code (a) Excavation Advance length [m] Face support 110 top heading 2 no 112 top heading 2 yes 150 top heading 4 no 152 top heading 4 yes 200 bench 2 no 202 bench 2 yes 220 bench 4 no 222 bench 4 yes G.H. Erharter et al.\n[page 6] Automation in Construction 127 (2021) 103701 6 highly influential on the stability of the excavation with longer advance lengths being more prone to failures than short ones (e.g. advance lengths in the Austrian standard ¨ONORM B2203–1 [46]: 1.0, 1.3, 1.7, 2.2, 3.0 and 4.0 m). After experimenting with advances lengths based on the standards in the beginning, we experienced that a big number of actions leads to a deterioration of the performance/confuses the agent in this model setup. This is also in accordance with other studies which have found that special measures/adaptations of the agent are necessary if the action space becomes increasingly complex [47,48]. The final set of actions in Table 3 therefore aims at giving the agent realistic options to choose from while also keeping the number of actions small (see the outlook in Section 7 for a discussion on increasing the number of actions). During an episode, the actions are chosen based on an “ε-greedy action selection process”, (see also “exploration vs. exploitation trade­ off”, e.g. [26]). ε is the exploration rate and is initially set to 1 (see “initial exploration in Table 2)."}], "source_title": "Reinforcement learning based process optimization and strategy development in conventional tunneling", "source_doi": "10.1016/j.autcon.2021.103701", "source_url": "https://www.sciencedirect.com/science/article/pii/S0926580521001527", "source_license": "cc-by", "zotero_item_key": "N5Q33JID"}
{"id": "daa4d90ab3b531d2", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Application of machine learning in predicting the rate-dependent compressive strength of rocks; DOI: 10.1016/j.jrmge.2022.01.008; license: cc-by-sa."}, {"role": "assistant", "content": "In recent years, with the development of computer and artificial intelligence technology, machine learning models have been widely employed to address the complexity of engineering issues, * Corresponding author. E-mail address: wu.wei@ntu.edu.sg (W. Wu). Peer review under responsibility of Institute of Rock and Soil Mechanics, Chi- nese Academy of Sciences. Contents lists available at ScienceDirect Journal of Rock Mechanics and Geotechnical Engineering journal homepage: www.jrmge.cn Journal of Rock Mechanics and Geotechnical Engineering 14 (2022) 1356e1365 https://doi.org/10.1016/j.jrmge.2022.01.008 1674-7755 \u0001 2022 Institute of Rock and Soil Mechanics, Chinese Academy of Sciences. Production and hosting by Elsevier B.V. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).\n[page 2] including the nonlinear behaviors of rocks. For example, the sup- port vector machine (SVM) algorithm was applied to forecast the static compressive strength and shear strength of rocks by input- ting rock density, point load strength, P-wave velocity, and slake durability index (Li and Tan, 2017). The back-propagation neural network (BPNN) algorithm was used to predict the static compressive strength and the elasticity modulus of rocks based on slake durability index, Schmidt hammer rebound number, effective porosity, and point load index (Yılmaz and Yuksek, 2008). Based on the random forest (RF) algorithm, the static compressive strength of rocks was estimated based on the available drilling data, such as weight on bit, drill string rotating speed, drilling torque, stand-pipe pressure, mud pumping rate, and penetration rate (Gamal et al., 2021). Given these successful applications to rock mechanics problems, the application of machine learning models can further be extended to rock dynamics to improve our ability to forecast the rate-dependent strength of rocks. This study aimed to use machine learning models to predict the compressive strength of rocks and to provide new insights into the rate-dependent strength of rocks. We first collected the laboratory data of the static compression tests and the SHPB tests on various types of hard rocks as well as the corresponding geometrical, physical, and mechanical properties. We then introduced the ma- chine learning algorithms (e.g. SVM, BPNN, and RF) and established three models based on these algorithms. We then trained these models using three quarters of the data and validated the models using the remaining data. Finally the importance of input param- eters and the effects of strain rate and specimen dimension on the compressive strength of rocks were discussed. 2. Data and methods 2.1. Data collection To train and validate the machine learning models in predicting the compressive strength of rocks, we first collected a few data sets of the static compression tests and the SHPB tests on eleven types of hard rocks from published papers (e.g. Frew et al., 2001; Li et al., 2005; Xia et al., 2008; Wang and Tonon, 2011; Zhang and Zhao, 2013; Zhou et al., 2017, 2020; Weng et al., 2020; Yu et al., 2021). Typical rock types in rock engineering were included, such as granite, marble, sandstone, and siltstone. The geometrical, physical, and mechanical properties of these rocks are listed in Table A1, including specimen length and diameter, grain size, bulk density, P- wave velocity, as well as compressive strength with the corre- sponding strain rate. The static compressive strength was consid- ered as rate-independent and approximated as a constant value, not specific for each specimen. Selection of the input parameters mainly relies on data availability, and other parameters, such as porosity, mineralogy, and microscale features (e.g. Fan et al., 2018, 2020), can also be used if available in the literature. To understand the simple correlation between any two pa- rameters in Table A1, we calculated the Pearson correlation coef- ficient (Edwards, 1976), a ratio of the covariance of two parameters to the product of their standard deviations. A correlation coefficient of 1 indicates a strong positive correlation, while \u00011 means a strong negative correlation. As illustrated in Fig. 1, for example, the P-wave velocity shows the strongest correlation with the static compres- sive strength, as they are highly related to the integrity of rocks. The P-wave velocity is also positively correlated with the grain size and the bulk density."}], "source_title": "Application of machine learning in predicting the rate-dependent compressive strength of rocks", "source_doi": "10.1016/j.jrmge.2022.01.008", "source_url": "https://www.sciencedirect.com/science/article/pii/S167477552200049X", "source_license": "cc-by-sa", "zotero_item_key": "KNCWQQJD"}
{"id": "5590c888ae9eb6bc", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: How to evaluate performance of prediction methods? Measures and their interpretation in variation effect analysis; DOI: 10.1186/1471-2164-13-s4-s2; license: cc-by."}, {"role": "assistant", "content": "Protein tolerance predictors Single nucleotide alterations are the most common genetic variation type. Human genomes contain these variations on average at every kilobase. Several computational meth- ods have been developed to classify these variations [1]. The evaluated methods were MutPred, nsSNPAnalyzer, Panther, PhD-SNP, PolyPhen, PolyPhen2, SIFT, SNAP, and SNPs&GO [5]. The methods differ in the properties of the variant they take into account, as well as in the nat- ure and the classification method. Panther, PhD-SNP and SIFT are based on evolutionary information. MutPred, nsSNPAnalyzer, PolyPhen2, SNAP and SNP&GO combine protein structural and/or functional parameters and sequence analysis derived information. Most of these are based on machine-learning methods. The positive test dataset included 19,335 missense var- iations from the PhenCode database [30], IDbases [31] and from 18 additional LSDBs. The negative dataset con- sisted of 21,170 nonsynonymous coding SNPs with an allele frequency >0.01 and chromosome sample count higher than 49 from the dbSNP database. As large num- bers of individual predictions were the Pathogenic-or-not Pipeline (PON-P) [32] was used for the submission of sequences and variants into the analysed programs. The performance was evaluated with the six measures described above. The performances of the programs ran- ged from poor (MCC 0.19) to reasonably good (MCC 0.65) [5]. It has been widely accepted that information about protein three dimensional structure would increase pre- diction performance. The very best methods use also structural and functional information, whereas others that are solely based on sequence level information per- form rather well. Further analyses were made to compare the methods pairwise, and to study whether the type of original or sub- stituting amino acid residue, the structural class of the protein, or the structural environment of the amino acid substitution, had an effect on the prediction performance. Existing programs thus have widely varying perfor- mance and there is still need for better methods. Con- sidering all the evaluation measures, no single method could be rated as best by all of them. Protein stability predictors Stability as a fundamental property affects protein func- tion, activity, and regulation. Changes to stability are often found to be involved in diseases. Systematic per- formance evaluation analysis has been made for eleven stability predictors performances including CUPSAT, Dmutant, FoldX, I-Mutant2.0, two versions of I- Mutant3.0 (sequence and structure versions), MultiMu- tate, MUpro, SCide, Scpred, and SRide [2]. SCide and Scpred, which predict stability centers, as well as SRide, which predicts stabilizing residues, predict only destabi- lizing effects, while all the others evaluate both stabiliz- ing and destabilizing changes. The major database for protein stability information is ProTherm [33]. The pruned dataset for testing con- tained 1784 variations from 80 proteins, with 1154 posi- tive cases of which 931 were destabilizing (ΔΔG ≥0.5 kcal/mol), 222 were stabilizing (ΔΔG ≤–0.5 kcal/mol), and 631 were neutral (0.5 kcal/mol ≥ΔΔG ≥–0.5 kcal/ mol). The majority of the methods had been trained using data from ProTherm, and thus only those cases that had been added to the database after training had occurred were used for testing. Of the measures recommended in here the authors used four, namely accuracy, specificity, sensitivity, and Vihinen BMC Genomics 2012, 13(Suppl 4):S2 http://www.biomedcentral.com/1471-2164/13/S4/S2 Page 8 of 10\n[page 9] MCC and the remaining row wise parameters could be calculated from the confusion tables. There were three groups of data, stability increasing, neutral and stability decreasing. The authors solved the problem of multiple classes by presenting three tables of results. The first one was grouped so that both stability increasing and decreasing were considered as pathogenic i.e. positive. In these analyses only two classes were con- sidered, stabilizing or destabilizing and neutral cases. The results for the all the cases show that accuracy ranges from 0.37 to 0.64 and MCC from -0.37 to only 0.12. All the programs succeeded better when predicting stability increasing or decreasing variations individually. The MCC reaches 0.35 and 0.38 for the methods best in predicting stability increasing and decreasing variants, respectively [2]."}], "source_title": "How to evaluate performance of prediction methods? Measures and their interpretation in variation effect analysis", "source_doi": "10.1186/1471-2164-13-s4-s2", "source_url": "https://bmcgenomics.biomedcentral.com/counter/pdf/10.1186/1471-2164-13-S4-S2", "source_license": "cc-by", "zotero_item_key": "LKJJ9WPZ"}
{"id": "cfcec27a84a2f614", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Spectral Angle Mapping and AI Methods Applied in Automatic Identification of Placer Deposit Magnetite Using Multispectral Camera Mounted on UAV; DOI: 10.3390/min12020268; license: cc-by-sa."}, {"role": "assistant", "content": "Consequently, this suggests that the Tree (Course–tree) model at 20 m performed better at discerning the targeted magnetite iron sand data compared to the other two UAV flight elevation model outputs.\n[page 14] Minerals 2022, 12, 268 14 of 19 Figure 9. Machine learning confusion matrices from Ensemble (RUS Boosted Trees) at 2 m UAV drone elevation, Tree (Course-tree) at 10 m UAV drone elevation, and Tree (Course-tree) at 20 m UAV drone. Reprinted with permission from ref. [1]. 2021 AusIMM [1] (p. 40). One more assessment that can be drawn from the Figure 9 confusion matrix is the average per-class precisions of 84.5%, 83.4% and 88.9% at 2 m, 10 m and 20 m flight elevations, respectively; at 20 m, the model continues to perform better. It can thus be said that the best classification ML model for this magnetite iron sands identification problem is Tree (Course tress) at 20 m. True Positive Rates (TPR): TPR = 100 \u0012 TP TP + FN \u0013 (2) False Negative Rates (FNR): FNR = 100 \u0012 FN TP + FN \u0013 (3) where FN is false negatives and TP is true positives. 3.3.2. Classification via Deep Learning CNN A convolution applies a filter over the multispectral data per band. Unlike with visual imagery data with height and width, the filters in a one dimensional CNN exhibit a 1D (spectral intensity) architecture instead of 2D as with normal visual images, hence the motivation to employ this type of CNN in this study as the datasets contain 6-bands deep spectral intensity data that require a 1D data processing method. The input dataset sizes for training and testing variables for each of the algorithms stood at 12,480,000 (from Figure 8) for each of the three flight elevations prior to obtaining the highest performing model. However, to train the 1D CNN, the spectral data were pre- processed in the following ratios: 70% (9,736,000) designated to training, 15% (1,872,000) validation and the remaining 15% (1,872,000) to testing. Moreover, all models were trained with Adam optimizer, which is said to have a best accuracy in enhancing CNN abilities in classification [19]. A learning rate of 0.001, a batch size of 128, and 25 epochs were also employed. As with ML, the 1D CNN training processes were conducted using the MATLAB\n[page 15] Minerals 2022, 12, 268 15 of 19 2020b deep learning toolbox with model selection at each flight elevation governed by global accuracy, average per class precision, and time taken to train each 1D CNN algorithm. At 2 m, 10 m and 20 m flight elevations, 80, 32 and 8 algorithms were trained and one model from the multitudes of trained models was selected. The same procedure was performed in training the ML models. Table 3 is a compilation of the best performing models at each of the three flight elevations. Results from the DL 1D CNN models in Table 3 show that the global accuracies as well as the average per-class precisions have an average of 99.8% and 99.3%, respectively. Moreover, based on global accuracy, the 1D CNN performs the same at 2 m and 10 m at 99.9%, followed by 20 m at 99.7%. However, the trend changes in considering the average per-class precisions where the best performing model is the 1 D CNN at 2 m (99.8%), followed by 20 m (99.4%) and 10 m (98.7%). Lastly, at 2 m, the 1D CNN takes a longer time to train (78 s) compared to the 1D CNNs at 10 m (45 s) and 20 m (68 s); this could suggest that the longer the training time, the higher the average per-class precision for this DL classification problem. Table 3. Deep learning one dimensional convolution neural network classification model comparisons at 2 m, 10 m and 20 m UAV flight elevations based 0.12, 0.13 and 0.17 SAM cosine similarity thresholds, respectively. Flight Height Global Accuracy (%) Average Per-Class Precision (%) Training Time (Seconds) 2 m 99.9% 99.8% 78 10 m 99.9% 98.7% 45 20 m 99.7% 99.4% 68 To further assess the viability of the best performing 1D CNN models for the three flight elevations, Figure 10 presents TPR and FNR performance metrics (Equations (2) and (3)). At 2 m, 10 m and 20 m, the 1D CNN models have a 100% TPR for magnetite iron sand (target). This suggests that the models are highly capable of deducing magnetite iron sands from the combination of magnetite iron sands (target) and non-magnetite beach sand (noise). On the other hand, the 1D CNN models attained a 99.6%, 97.4% and 98.7% TPR for non-magnetite beach sand (noise) at 2 m, 10 m and 20 m, respectively, which is considerably high."}], "source_title": "Spectral Angle Mapping and AI Methods Applied in Automatic Identification of Placer Deposit Magnetite Using Multispectral Camera Mounted on UAV", "source_doi": "10.3390/min12020268", "source_url": "https://www.mdpi.com/2075-163X/12/2/268/pdf?version=1645347690", "source_license": "cc-by-sa", "zotero_item_key": "VPIE7856"}
{"id": "e8271768f3eecee8", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: How to evaluate performance of prediction methods? Measures and their interpretation in variation effect analysis; DOI: 10.1186/1471-2164-13-s4-s2; license: cc-by."}, {"role": "assistant", "content": "[page 1] PROCEEDINGS Open Access How to evaluate performance of prediction methods? Measures and their interpretation in variation effect analysis Mauno Vihinen1,2,3 From SNP-SIG 2011: Identification and annotation of SNPs in the context of structure, function and disease Vienna, Austria. 15 July 2011 Abstract Background: Prediction methods are increasingly used in biosciences to forecast diverse features and characteristics. Binary two-state classifiers are the most common applications. They are usually based on machine learning approaches. For the end user it is often problematic to evaluate the true performance and applicability of computational tools as some knowledge about computer science and statistics would be needed. Results: Instructions are given on how to interpret and compare method evaluation results. For systematic method performance analysis is needed established benchmark datasets which contain cases with known outcome, and suitable evaluation measures. The criteria for benchmark datasets are discussed along with their implementation in VariBench, benchmark database for variations. There is no single measure that alone could describe all the aspects of method performance. Predictions of genetic variation effects on DNA, RNA and protein level are important as information about variants can be produced much faster than their disease relevance can be experimentally verified. Therefore numerous prediction tools have been developed, however, systematic analyses of their performance and comparison have just started to emerge. Conclusions: The end users of prediction tools should be able to understand how evaluation is done and how to interpret the results. Six main performance evaluation measures are introduced. These include sensitivity, specificity, positive predictive value, negative predictive value, accuracy and Matthews correlation coefficient. Together with receiver operating characteristics (ROC) analysis they provide a good picture about the performance of methods and allow their objective and quantitative comparison. A checklist of items to look at is provided. Comparisons of methods for missense variant tolerance, protein stability changes due to amino acid substitutions, and effects of variations on mRNA splicing are presented. Background Gene and genome sequencing speed is ever increasing and thus lots of genetic variation information is available. The technological development of sequencing methods has led to a situation where the interpretation of the gen- erated data is a severe bottleneck for the use of genetic information. Numerous prediction methods have been developed during the last decade to address the relevance of gene and protein variants to pathogenicity. General tolerance methods predict whether the variants are dis- ease-related or not (or affect protein function or not), and specific methods are used to address variation effect mechanisms [1]. These methods can be useful. However, until recently their true applicability and performance have not been studied systematically [2-5]. When meth- ods are originally published, authors provide some performance measures, which are usually not comparable with other methods due to the use of different training and test datasets, different reported measures etc. The scope of this article is to discuss how the assessment of method performance should be done and interpretation Correspondence: mauno.vihinen@med.lu.se 1Institute of Biomedical Technology, FI-33014 University of Tampere, Finland Full list of author information is available at the end of the article Vihinen BMC Genomics 2012, 13(Suppl 4):S2 http://www.biomedcentral.com/1471-2164/13/S4/S2 © 2012 Vihinen; licensee BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.\n[page 2] of the results and the choice of the best methods. The text is mainly intended for scientists who are users of predictors without training in statistics or computer science. Method developers are taken into account by providing a checklist of items to be reported with meth- ods. The examples discussed are related to prediction of variant effects, but description of methods and evaluation measures is general and thereby not application domain specific."}], "source_title": "How to evaluate performance of prediction methods? Measures and their interpretation in variation effect analysis", "source_doi": "10.1186/1471-2164-13-s4-s2", "source_url": "https://bmcgenomics.biomedcentral.com/counter/pdf/10.1186/1471-2164-13-S4-S2", "source_license": "cc-by", "zotero_item_key": "LKJJ9WPZ"}
{"id": "f6df0f2ceeb590fc", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Automated Excavator Based on Reinforcement Learning and Multibody System Dynamics; DOI: 10.1109/access.2020.3040246; license: cc-by-sa."}, {"role": "assistant", "content": "García, and R. J. Martínez, ‘‘Approximation of con- tinuous media models for granular systems using cellular automata,’’ in Cellular Automata (Lecture Notes in Computer Science), vol. 3305. Berlin, Germany: Springer, 2004, pp. 230–237, doi: 10.1007/978-3-540-30479- 1_24. [23] M. Pla-Castells, I. García-Fernández, and R. J. Martínez, ‘‘Interactive terrain simulation and force distribution models in sand piles,’’ in Cellular Automata, vol. 4173. Berlin, Germany: Springer, 2006, pp. 392–401, doi: 10.1007/11861201_46. [24] D. Holz, T. Beer, and T. Kuhlen, ‘‘Soil deformation models for real- time simulation: A hybrid approach,’’ in Proc. Vriphys 6th Workshop Virtual Real. Interact. Phys. Simulations, Jan. 2009, pp. 21–30, doi: 10.2312/PE/vriphys/vriphys09/021-030. [25] S. Jaiswal, P. Korkealaakso, R. Aman, J. Sopanen, and A. Mikkola, ‘‘Deformable terrain model for the real-time multibody simulation of a tractor with a hydraulically driven front-loader,’’ IEEE Access, vol. 7, pp. 172694–172708, 2019, doi: 10.1109/ACCESS.2019.2956164. [26] P. Hintjens, ‘‘ZeroMQ,’’ O’Reilly Media, Sebastopol, CA, USA, Tech. Rep., 2013, vol. 53, no. 9. [27] Part 2: Kinds of RL Algorithms—Spinning Up documentation. Accessed: May 29, 2020. [Online]. Available: https://spinningup.openai.com/en/ latest/spinningup/rl_intro2.html [28] J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, ‘‘Prox- imal policy optimization algorithms,’’ 2017, arXiv:1707.06347. [Online]. Available: http://arxiv.org/abs/1707.06347 ILYA KURINOV received the B.S. degree in mechanical engineering from the Saimaa Univer- sity of Applied Sciences and the M.S. degree in mechatronics from the Lappeenranta University of Technology. Since 2018, he has been working as a Ph.D. Researcher with the Machine Design Labo- ratory, LUT University. His research inter- ests include application of multibody system dynamics and machine learning algorithms in automation of mechatronic machines. GRZEGORZ ORZECHOWSKI received the B.S. and M.S. degrees and the Ph.D. degree in automa- tion and robotics from the Warsaw University of Technology, Warsaw, Poland, in 2007 and 2012. From 2012 to 2017, he was an Assistant Profes- sor with the Division of Theory of Machines and Robots, Warsaw University of Technology. Since 2017, he has been a Postdoctoral Researcher with the Laboratory of Machine Design, LUT Univer- sity, Lappeenranta, Finland. His research interests include the computational methods in mechanics, studies of the deformable models in challenging dynamical excitations, and reinforcement learning techniques in application to machine control. PERTTU HÄMÄLÄINEN received the M.Sc. (Tech.) degree from the Helsinki University of Technology, in 2001, the M.A. degree in new media from the University of Art and Design, Helsinki, in 2002, and the Ph.D. degree in com- puter science from the Helsinki University of Technology, in 2007. He is currently an Associate Professor with Aalto University. He has published widely on human–computer interaction, computer animation and movement synthesis, and game research. AKI MIKKOLA received the Ph.D. degree in the field of machine design in 1997. Since 2002, he has been working as a Pro- fessor with the Department of Mechanical Engi- neering, Lappeenranta University of Technology, Lappeenranta, Finland. He is currently leading the Research Team of the Laboratory of Machine Design, Lappeenranta University of Technology. His research interests include machine dynamics and vibration, multibody system dynamics, and bio-mechanics. He has been awarded five patents and has contributed to more than 90 peer-reviewed journal articles. 214006 VOLUME 8, 2020"}], "source_title": "Automated Excavator Based on Reinforcement Learning and Multibody System Dynamics", "source_doi": "10.1109/access.2020.3040246", "source_url": "https://ieeexplore.ieee.org/ielx7/6287639/6514899/09268069.pdf", "source_license": "cc-by-sa", "zotero_item_key": "7F2IQNFD"}
{"id": "d00e1bc704a00341", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Two-phase multi-model automatic brain tumour diagnosis system from magnetic resonance images using convolutional neural networks; DOI: 10.1186/s13640-018-0332-4; license: cc-by-sa."}, {"role": "assistant", "content": "[page 1] EURASIP Journal on Image and Video Processing Abd-Ellah et al. EURASIP Journal on Image and Video Processing (2018) 2018:97 https://doi.org/10.1186/s13640-018-0332-4 RESEARCH Open Access Two-phase multi-model automatic brain tumour diagnosis system from magnetic resonance images using convolutional neural networks Mahmoud Khaled Abd-Ellah1, Ali Ismail Awad2,3* , Ashraf A. M. Khalaf4 and Hesham F. A. Hamed4 Abstract Brain tumour is a serious disease, and the number of people who are dying due to brain tumours is increasing. Manual tumour diagnosis from magnetic resonance images (MRIs) is a time consuming process and is insufficient for accurately detecting, localizing, and classifying the tumour type. This research proposes a novel two-phase multi-model automatic diagnosis system for brain tumour detection and localization. In the first phase, the system structure consists of preprocessing, feature extraction using a convolutional neural network (CNN), and feature classification using the error-correcting output codes support vector machine (ECOC-SVM) approach. The purpose of the first system phase is to detect brain tumour by classifying the MRIs into normal and abnormal images. The aim of the second system phase is to localize the tumour within the abnormal MRIs using a fully designed five-layer region-based convolutional neural network (R-CNN). The performance of the first phase was assessed using three CNN models, namely, AlexNet, Visual Geometry Group (VGG)-16, and VGG-19, and a maximum detection accuracy of 99.55% was achieved with AlexNet using 349 images extracted from the standard Reference Image Database to Evaluate Response (RIDER) Neuro MRI database. The brain tumour localization phase was evaluated using 804 3D MRIs from the Brain Tumor Segmentation (BraTS) 2013 database, and a DICE score of 0.87 was achieved. The empirical work proved the outstanding performance of the proposed deep learning-based system in tumour detection compared to other non-deep-learning approaches in the literature. The obtained results also demonstrate the superiority of the proposed system concerning both tumour detection and localization. Keywords: Brain tumour diagnosis, MRI segmentation, Tumour detection and localization, Convolutional neural networks (CNNs) 1 Introduction Brain tumour is a serious disease in which an abnormal growth of tissue inside the the brain can disrupt proper brain function. The National Brain Tumor Foundation (NBTF) reported that the number of people in devel- oped countries who have died due to brain tumours has increased by 300% over the last three decades [1, 2]. Man- ual brain tumour diagnosis by physicians is a less accurate yet still time-consuming procedure [3, 4]. *Correspondence: ali.awad@ltu.se 2Department of Computer Science, Electrical and Space Engineering, Luleå University of Technology, SE-97187 Luleå, Sweden 3Faculty of Engineering, Al-Azhar University, P.O. Box 83513, Qena, Egypt Full list of author information is available at the end of the article Magnetic resonance imaging is an advanced medical imaging technique providing rich information about the human soft tissue anatomy [5]. Automatic brain tumour detection from magnetic resonance images (MRIs) aims to classify MRIs into normal and abnormal according to the absence or presence of a tumour, respectively. Thus, the diagnosis problem can be addressed as an image clas- sification challenge. The pervasiveness of brain tumours has led to the production of a massive amount of MRI data. Therefore, developing an automatic brain tumour diagnosis system that achieves high tumour detection and localization accuracies is a vital need [6]. Machine learn- ing in general and deep learning approaches in particular play core roles in computer-assisted brain image analysis, © The Author(s). 2018 Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.\n[page 2] Abd-Ellah et al. EURASIP Journal on Image and Video Processing (2018) 2018:97 Page 2 of 10 segmentation, registration, and tumour tissue classifica- tion [7, 8]."}], "source_title": "Two-phase multi-model automatic brain tumour diagnosis system from magnetic resonance images using convolutional neural networks", "source_doi": "10.1186/s13640-018-0332-4", "source_url": "https://jivp-eurasipjournals.springeropen.com/track/pdf/10.1186/s13640-018-0332-4", "source_license": "cc-by-sa", "zotero_item_key": "2ZZSU54R"}
{"id": "9ec911861e72fca1", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Intelligent Classification of Surrounding Rock of Tunnel Based on 10 Machine Learning Algorithms; DOI: 10.3390/app12052656; license: cc-by-sa."}, {"role": "assistant", "content": "2022, 12, 2656 10 of 20 (a) (b) (c) Figure 7. Scatter diagram of the data set after over-sampling (1260). (a) Drilling parameters of surround rock grade-III (420); (b) drilling parameters of surround rock grade-IV (420); (c) drilling parameters of surround rock grade-V (420). 2.5. Sample Feature Analysis Figure 6 shows that all median, mean, and upper and lower quartile values of each drilling parameter in the box diagram monotonically change with the surrounding rock grade, which preliminarily indicates a correlation between the drilling parameters and the surrounding rock grade.\n[page 11] Appl. Sci. 2022, 12, 2656 11 of 20 Pearson correlation coefficient [56] was used to analyze the correlations between the surrounding rock grade and the drilling parameters in 1260 samples, and the correlation coefficients are shown in Table 5. Table 5. Correlation coefficients between the surrounding rock grade and each drilling parameter. Index Vp Ph Pf Pr R 0.59 −0.64 −0.52 −0.65 In Table 5, all the absolute value of the correlation coefficients between the four drilling parameters and the surrounding rock grades are above 0.5, which is a weak correlation, and the correlation coefficients are close to each other. Therefore, these four indices were selected for the surrounding rock classification. 2.6. Sample Data Splitting According to the above 1260 data samples, the intelligent surrounding rock classifica- tion sample database was established. In order that the selected parameters can represent the majority of samples, the majority samples are selected for training and the minority samples for prediction in the parameter optimization stage (validation process). As to the model evaluation stage (testing process), we selected more samples for prediction than the previous stage so that the performance evaluation of the model is more convincingly. Thus, the sample database after random sequencing was divided into the training set and prediction set using the five-fold cross verification method in the validation process, and as to the testing process, the ratio is three-fold. The specific distribution is shown in Table 6. Table 6. Intelligent surrounding rock classification sample database distribution. Item Surrounding Rock III IV V Total Validation process Training set 336 336 336 1008 Prediction set (Validation set) 84 84 84 252 Testing process Training set 280 280 280 840 Prediction set (Testing set) 140 140 140 420 In the field of machine learning, different evaluation indexes (that is, different features in feature vectors are described as different evaluation indexes) often have different di- mensional and dimensional units, which will affect the results of data analysis. In order to eliminate the dimensional influence between indexes, data normalization is required. Normalization means that the input data are limited to a certain range, and this time the data are limited 0 to 1. The data are normalized using the min–max normalization method, as defined in Equation (2). x′ = x −xmin xmax −xmin (2) where x is the value of the original data, xmax is the max value of the original data, xmin is the min value of the original data, x’ is the normalized value. The normalization needs to be carried out in the training set in Table 6 first, and then the normalization of the prediction set is completed by taking the normalization parameters (xmax and xmin) of the training set.\n[page 12] Appl. Sci. 2022, 12, 2656 12 of 20 2.7. Model Parameter Optimization and Evaluation Based on the python machine learning toolkit (Scikit-learn), 10 types of supervised ma- chine learning algorithms were used to train the intelligent surrounding rock classification model: the support vector machine (SVM), back propagation neural network (BP) [55,56], radial basis function neural network (RBF), K-nearest neighbor (KNN), Gaussian naive Bayes (GNB), decision tree (DT), random forest (RF), extra trees (ET), bootstrap aggregating (Bagging) and gradient boosting (GB) algorithms. The model parameters were optimized and evaluated by the grid search K-flod cross validation method (k = 5), and the model performance was externally unbiased when using the method of K-flod cross validation (k = 3). In the model performance optimization process, the average accuracy was adopted to determine the optimum hyperparameter combinations."}], "source_title": "Intelligent Classification of Surrounding Rock of Tunnel Based on 10 Machine Learning Algorithms", "source_doi": "10.3390/app12052656", "source_url": "https://www.mdpi.com/2076-3417/12/5/2656/pdf?version=1646383280", "source_license": "cc-by-sa", "zotero_item_key": "UGAERXYU"}
{"id": "8ab15d3d71c6d683", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Detection of Ionospheric Scintillation Based on XGBoost Model Improved by SMOTE-ENN Technique; DOI: 10.3390/rs13132577; license: cc-by-sa."}, {"role": "assistant", "content": "Testing Set 1 2 3 4 5 6 7 Accuracy (%) Decision tree 2.58 1.57 2.57 0.47 1.18 0.44 1.06 Random forest 2.45 2.07 2.34 0.65 1.24 0.36 0.45 XGBoost 6.04 2.44 2.92 0.94 0.62 1.25 0.29 Recall for class 2 (%) Decision tree 4.09 2.58 10.94 9.15 9.24 8.01 8.41 Random forest 5.82 9.23 10.61 4.51 13.90 7.00 12.61 XGBoost 10.59 12.60 17.69 13.98 16.17 19.40 14.06 F1 score for class 2 (%) Decision tree 2.28 1.76 6.53 2.76 5.46 1.06 4.51 Random forest 3.17 5.01 5.62 1.29 7.5 2.11 5.37 XGBoost 6.00 6.64 9.54 7.15 8.04 10.61 6.61 Figure 10. Cont.\n[page 17] Remote Sens. 2021, 13, 2577 17 of 22 Figure 10. Testing results for detection algorithms regarding overall accuracy (top row), recall value for class 2 (middle row) and F1 score for class 2 (bottom row), with different ratios of classes 0/1/2 for testing datasets. Relevant information on training dataset and seven testing datasets is listed in Table 7. The green, yellow and orange bars refer to results for decision tree, random forest and XGBoost, respectively. The highlighted bold bar on the right of each group represents the result of the XGBoost method improved by the SMOTE-ENN technique, with better performance. 5. Discussion All the experiments were carried out on real data collected in an equatorial region during solar maximum years, under various ionospheric scintillation intensities. Firstly, different resampling techniques were compared based on the XGBoost algorithm, and the integration of SMOTE and ENN produced better results regarding XGBoost’s performance. As shown in Table 4, based on random oversampling, there was only a 3% increase in the recall value for the minority class 2 in that it just randomly took out samples in class 2 without any rule, and thus added in several invalid data. Meanwhile, there was little contribution to the other two classes. Different from oversampling, the method of random undersampling is aimed at randomly deleting samples of the majority class until the data of these classes are equal to those of the minority classes in number. Thus, much useful information is discarded, reducing the detection accuracy, presenting similar results to random oversampling. Although the resampling rule of SMOTE is based on the nearest neighboring samples and linear interpolation, it still aims at increasing the minority class\n[page 18] Remote Sens. 2021, 13, 2577 18 of 22 without significant improvement for the other two classes. As for the SMOTE, ENN only showed a 4% improvement on the recall value of the majority class 0, because the samples of the boundary around classes 0 and 1 were deleted. Meanwhile, the overall samples were still imbalanced, resulting in ineffective enhancement for class 2. Differently, the SMOTE-ENN method firstly deals with the data of the minority class through the SMOTE oversampling technique and then carries out the ENN undersampling for the data of the majority class. It not only increases the data of the minority class but also deletes the multi-class samples on the boundary as much as possible, significantly improving the detection performance in various aspects. The results in Table 4 show that there was an apparent improvement in performance for all three classes, with an approximately 3% increase in overall accuracy. As for the value of recall, it showed a 5% enhancement for class 2, corresponding to strong scintillation events, and 4% increase for class 0, corresponding to weak scintillation events, meaning that more strong and weak scintillation events that may be easily missed can be detected correctly. Thereby, the precision for class 1, corresponding to moderate scintillation events, increased by 6%, with more events of classes 0 and 2 being correctly detected. Meanwhile, the F1 scores for the three classes increased by 3%, 3% and 2%, respectively, demonstrating an improvement in both precision and recall. Especially, the integrated technique outperforms ENN by 4.13% in recall for class 2, while showing a 1.37% enhancement in overall accuracy compared with SMOTE. Compared with the other single resampling methods, the SMOTE-ENN proved to be effective in dealing with imbalanced data with insufficient strong scintillation events. Based on that, brief comparisons of five methods were drawn."}], "source_title": "Detection of Ionospheric Scintillation Based on XGBoost Model Improved by SMOTE-ENN Technique", "source_doi": "10.3390/rs13132577", "source_url": "https://www.mdpi.com/2072-4292/13/13/2577/pdf?version=1625212202", "source_license": "cc-by-sa", "zotero_item_key": "4YEIWMK3"}
{"id": "acf529a5684722a1", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Reinforcement learning based process optimization and strategy development in conventional tunneling; DOI: 10.1016/j.autcon.2021.103701; license: cc-by."}, {"role": "assistant", "content": "Winkler, Machine learning in tunnelling – capabilities and challenges, Geomechanik Tunnelbau 13 (2020) 191–198, https:// doi.org/10.1002/geot.202000001. [2] DAUB, BIM in Tunnelling: Digital Design, Building and Operation of Underground Structures. http://www.daub-ita.de/fileadmin/documents/daub/gtcrec4/gtcre c11v3_BIM_in_Tunnelling_05-2019.pdf, 2019 (accessed 9 October 2020). [3] L. Wagner, A. Kluckner, C.M. Monsberger, P. Wolf, K. Prall, W. Schubert, W. Lienhart, Direct and distributed strain measurements inside a Shotcrete lining: concept and realisation, Rock Mech. Rock. Eng. 53 (2020) 641–652, https://doi. org/10.1007/s00603-019-01923-4. [4] G.H. Erharter, T. Marcher, C. Reinhold, Application of artificial neural networks for underground construction – chances and challenges – insights from the BBT exploratory tunnel Ahrental Pfons, Geomechanik Tunnelbau 12 (2019) 472–477, https://doi.org/10.1002/geot.201900027. [5] G.H. Erharter, T. Marcher, C. Reinhold, Comparison of artificial neural networks for TBM data classification, in: Proceedings of the 14th International Congress on Rock Mechanics and Rock Engineering (ISRM 2019), Foz de Iguassu, Brazil, 2019. [6] G.H. Erharter, T. Marcher, C. Reinhold, Artificial neural network based online rockmass behavior classification of TBM data, in: Information Technology in Geo- Engineering, firstst ed. twentiethtwentieth, Springer, 2020, pp. 178–188. [7] R. Wu, Y. Fujita, K. Soga, Integrating domain knowledge with deep learning models: an interpretable AI system for automatic work progress identification of NATM tunnels, Tunn. Undergr. Space Technol. 105 (2020) 103558, https://doi. org/10.1016/j.tust.2020.103558. [8] S.K. Shreyas, A. Dey, Application of soft computing techniques in tunnelling and underground excavations: state of the art and future prospects, Innov. Infrastruct. Solut. 4 (2019), https://doi.org/10.1007/s41062-019-0234-z. [9] W. Zhang, R. Zhang, C. Wu, A.T.C. Goh, S. Lacasse, Z. Liu, H. Liu, State-of-the-art review of soft computing applications in underground excavations, Geosci. Front. 11 (2020) 1095–1106, https://doi.org/10.1016/j.gsf.2019.12.003. [10] S. Isam, Z. Wengang, Use of soft computing techniques for tunneling optimization of tunnel boring machines, Underground Space (2020), https://doi.org/10.1016/j. undsp.2019.12.001. [11] B.B. Sheil, S.K. Suryasentana, M.A. Mooney, H. Zhu, Machine learning to inform tunnelling operations: recent advances and future trends, Proc. Inst. Civ. Eng. (2020) 1–18, https://doi.org/10.1680/jsmic.20.00011. [12] Q. Zhang, Z. Liu, J. Tan, Prediction of geological conditions for a tunnel boring machine using big operational data, Autom. Constr. 100 (2019) 73–83, https://doi. org/10.1016/j.autcon.2018.12.022. [13] G.H. Erharter, T. Marcher, MSAC: towards data driven system behavior classification for TBM tunneling, Tunn. Undergr. Space Technol. 103 (2020) 103466, https://doi.org/10.1016/j.tust.2020.103466. [14] V. Mnih, K. Kavukcuoglu, D. Silver, A.A. Rusu, J. Veness, M.G. Bellemare, A. Graves, M. Riedmiller, A.K. Fidjeland, G. Ostrovski, S. Petersen, C. Beattie, A. Sadik, I. Antonoglou, H. King, D. Kumaran, D. Wierstra, S. Legg, D. Hassabis, Human-level control through deep reinforcement learning, Nature 518 (2015) 529–533, https://doi.org/10.1038/nature14236. [15] D. Silver, A. Huang, C.J. Maddison, A. Guez, L. Sifre, G. van den Driessche, J. Schrittwieser, I. Antonoglou, V. Panneershelvam, M. Lanctot, S. Dieleman, D. Grewe, J. Nham, N. Kalchbrenner, I. Sutskever, T. Lillicrap, M. Leach, K. Kavukcuoglu, T. Graepel, D. Hassabis, Mastering the game of go with deep Fig. 11. Exemplary log of one episode. The x-axis shows the number of blasts/moves the agent needed to achieve breakthrough (i.e. 105): Top row: Time-distance diagram showing the logged position of top heading excavation (solid line), bench excavation (dashed line) and the difference in between (grey area). Middle row: Record of the actions (see Table 3 for the corresponding action codes). Bottom row: cumulative reward throughout the episode. G.H. Erharter et al.\n[page 12] Automation in Construction 127 (2021) 103701 12 neural networks and tree search, Nature 529 (2016) 484–489, https://doi.org/ 10.1038/nature16961. [16] J. Shahrabi, M.A. Adibi, M. Mahootchi, A reinforcement learning approach to parameter estimation in dynamic job shop scheduling, Comput. Ind. Eng. 110 (2017) 75–82, https://doi.org/10.1016/j.cie.2017.05.026. [17] M."}], "source_title": "Reinforcement learning based process optimization and strategy development in conventional tunneling", "source_doi": "10.1016/j.autcon.2021.103701", "source_url": "https://www.sciencedirect.com/science/article/pii/S0926580521001527", "source_license": "cc-by", "zotero_item_key": "RXHUVV3X"}
{"id": "bc24413b456db9b6", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Detection of Ionospheric Scintillation Based on XGBoost Model Improved by SMOTE-ENN Technique; DOI: 10.3390/rs13132577; license: cc-by-sa."}, {"role": "assistant", "content": "As shown in Table 5, XGBoost outperformed the CNN by 1.57%, 12.92% and 8.99% in terms of the overall accuracy, recall and F1 score for class 2 when trained and tested on raw data, showing significant superiority in strong scintillation detection. Overall, all four detection algorithms were enhanced when improved by SMOTE-ENN. Moreover, for resampled data, XGBoost outperformed the SVM by 0.65% in terms of overall accuracy. The recall for the SVM was 3.46% higher than that for XGBoost, but the F1 score for the SVM was 4.20% lower. This indicates that the SVM missed many strong scintillation events, with a precision of only 83.79% for class 2. More importantly, larger computational loads and longer running times are needed for SVM detection than detection with the three other machine learning methods. Thus, both methods will not be further discussed due to the unavailability of the CNN and large running loads for the SVM. For the different imbalanced training sets, the accuracy of the improved XGBoost was about 4% to 5% higher than that of the decision tree and random forest, which is clear from Figure 9. Although the recall values of the decision tree were about 2% higher than those of XGBoost based on the training sets of 1, 3 and 6, the F1 scores for XGBoost were about 2% higher than those for the decision tree and random forest. As shown in the right panels of Figure 9, there was a 6.39% decline in the recall value, from 90.86% to 84.47%, and 2.82% decline in the F1 score, from 92.81% to 89.99%, implying more missed strong scintillation events. However, the XGBoost model, trained on the SMOTE-ENN resampled datasets, remained relatively stable, with recall values of about 90% to 91% and F1 scores of 92% to 93%, indicating more enhancement for more severely imbalanced training samples. Meanwhile, there was a 6.37% improvement in the recall value, from 84.47% to 90.84%, and 2.78% improvement in the F1 score, from 89.99% to 92.77%, with the resampled training data with a sample ratio of 7:7:1. Additionally, there was also improvement in the other two classes in the precision, recall and F1 score, as well as the overall accuracy, which increased by 3% to over 92% compared to that with the raw data. These illustrate the effectiveness of the proposed XGBoost algorithm improved by the SMOTE-ENN resampling technique for imbalanced training data in improving the detection performance for strong scintillation events. Considering various natural phenomena of different ionospheric scintillation events, seven groups of testing datasets with different degrees of imbalance were established, and each dataset included nine segments of real one-hour data. The results significantly illus-\n[page 19] Remote Sens. 2021, 13, 2577 19 of 22 trate the excellent performance of the improved XGBoost in strong scintillation detection under different scintillation conditions. It can be intuitively observed from Figure 10 that the accuracy of the improved XGBoost was higher than that of the other methods based on each testing dataset, while almost all the values of recall and F1 score were similar. The overall testing accuracy of the XGBoost model trained on the resampled dataset varied from 90.42% to 96.04%, higher than that of the XGBoost model trained on the raw dataset, ranging from 85.81% to 95.42%. Table 8 illustrates the improvement ratios for the evaluation indicators for the different testing sets with different detection algorithms trained on the resampled dataset, compared with the raw dataset. For these testing datasets, the results show an increase of at least 10% in the recall values. Especially, for the sixth group of the testing dataset, it increased by 19.40%, from 74.61% to 94.01%. Meanwhile, the F1 score significantly increased by more than 6.00%, ranging from 77.31% to 94.82%, for these seven testing datasets, while there was a significant increase of 10.61% from 84.14% to 94.75% for the sixth testing dataset. The results also present an improvement or maintenance of the corresponding recall, precision and F1 scores for the other two classes. These results indicate that it is valuable to enhance the detection accuracy for strong scintillation events with different degrees of imbalance in the testing data with the method of resampling the imbalanced training data by SMOTE-ENN before training the XGBoost model. 6."}], "source_title": "Detection of Ionospheric Scintillation Based on XGBoost Model Improved by SMOTE-ENN Technique", "source_doi": "10.3390/rs13132577", "source_url": "https://www.mdpi.com/2072-4292/13/13/2577/pdf?version=1625212202", "source_license": "cc-by-sa", "zotero_item_key": "4YEIWMK3"}
{"id": "d7ac918e129ab17f", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Detection of Ionospheric Scintillation Based on XGBoost Model Improved by SMOTE-ENN Technique; DOI: 10.3390/rs13132577; license: cc-by-sa."}, {"role": "assistant", "content": "IEEE Trans. Knowl. Data Eng. 2009, 21, 1263–1284.\n[page 22] Remote Sens. 2021, 13, 2577 22 of 22 43. Wilson, D.L. Asymptotic properties of nearest neighbor rules using edited data. IEEE Trans. Syst. Man Cybern. 1972, 2, 408–421. [CrossRef] 44. Tomek, I. Two modifications of CNN. IEEE Trans. Syst. Man Cybern. 1976, 6, 769–772. 45. Hart, P.E. The condensed nearest neighbor rule. IEEE Trans. Inf. Theory 1967, 14, 515–516. [CrossRef] 46. Laurikkala, J. Improving identification of difficult small classes by balancing class distribution. In Proceedings of the 8th Conference on Artificial Intelligence in Medicine in Europe, Cascais, Portugal, 1–4 July 2001; pp. 63–66. 47. Chawla, N.V.; Bowyer, K.W.; Hall, L.O.; Kegelmeyer, W.P. SMOTE: Synthetic minority over-sampling technique nitesh. J. Artif. Intell. Res. 2002, 16, 321–357. [CrossRef] 48. Han, H.; Wang, W.; Mao, B. Borderline-SMOTE: A new over-sampling method in imbalanced data sets learning. In Proceedings of the Advances in Intelligent Computing—International Conference on Intelligent Computing, Hefei, China, 23–26 August 2005; pp. 878–887. 49. Batista, G.E.A.P.A.; Prati, R.C.; Monard, M.C. A study of the behavior of several methods for balancing machine learning training data. ACM SIGKDD Explor. Newsl. 2004, 6, 20–29. [CrossRef]"}], "source_title": "Detection of Ionospheric Scintillation Based on XGBoost Model Improved by SMOTE-ENN Technique", "source_doi": "10.3390/rs13132577", "source_url": "https://www.mdpi.com/2072-4292/13/13/2577/pdf?version=1625212202", "source_license": "cc-by-sa", "zotero_item_key": "4YEIWMK3"}
{"id": "25b841cc37b4da67", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Application of machine learning in predicting the rate-dependent compressive strength of rocks; DOI: 10.1016/j.jrmge.2022.01.008; license: cc-by-sa."}, {"role": "assistant", "content": "[page 1] Full Length Article Application of machine learning in predicting the rate-dependent compressive strength of rocks Mingdong Wei a, Wenzhao Meng a, Feng Dai b, Wei Wu a,* a School of Civil and Environmental Engineering, Nanyang Technological University, Singapore b State Key Laboratory of Hydraulics and Mountain River Engineering, College of Water Resource and Hydropower, Sichuan University, Chengdu, 610065, China a r t i c l e i n f o Article history: Received 26 October 2021 Received in revised form 15 December 2021 Accepted 20 January 2022 Available online 18 March 2022 Keywords: Machine learning Rock dynamics Compressive strength Strain rate a b s t r a c t Accurate prediction of compressive strength of rocks relies on the rate-dependent behaviors of rocks, and correlation among the geometrical, physical, and mechanical properties of rocks. However, these prop- erties may not be easy to control in laboratory experiments, particularly in dynamic compression ex- periments. By training three machine learning models based on the support vector machine (SVM), back- propagation neural network (BPNN), and random forest (RF) algorithms, we isolated different input parameters, such as static compressive strength, P-wave velocity, specimen dimension, grain size, bulk density, and strain rate, to identify their importance in the strength prediction. Our results demonstrated that the RF algorithm shows a better performance than the other two algorithms. The strain rate is a key input parameter influencing the performance of these models, while the others (e.g. static compressive strength and P-wave velocity) are less important as their roles can be compensated by alternative pa- rameters. The results also revealed that the effect of specimen dimension on the rock strength can be overshadowed at high strain rates, while the effect on the dynamic increase factor (i.e. the ratio of dy- namic to static compressive strength) becomes significant. The dynamic increase factors for different specimen dimensions bifurcate when the strain rate reaches a relatively high value, a clue to improve our understanding of the transitional behaviors of rocks from low to high strain rates. \u0001 2022 Institute of Rock and Soil Mechanics, Chinese Academy of Sciences. Production and hosting by Elsevier B.V. This is an open access article under the CC BY license (http://creativecommons.org/licenses/ by/4.0/). 1. Introduction Rate dependence of rock strength is a critical issue in rock dy- namics, as rock behaviors vary drastically at different strain rates. Understanding the rate-dependent strengths of rocks is important for rock engineering design and construction (Zhao et al., 1999; Hudson, 2012). Our current approaches in determining the compressive strength of rocks rely basically on laboratory experi- ments, such as the static compression test (ISRM, 1979) and the split Hopkinson pressure bar (SHPB) test (Zhou et al., 2012). These test methods investigate the rate dependence of compressive strength of rocks at low and high strain rates, respectively. A gen- eral understanding is that the compressive strength slightly in- creases with a higher strain rate at a low strain rate (e.g. < 5 s\u00011) and grows rapidly at a high strain rate (e.g. > 500 s\u00011). The test methods are difficult to be used directly to investigate the compressive strength of rocks at intermediate strain rates (5e500 s\u00011), which are of interest to understand the mechanisms of excavation-induced geohazards such as rockburst (Duan et al., 2019a; Zhong et al., 2021). Efforts have been made to modify the test setups to control the intermediate strain rates (e.g. Li et al., 2005; Whittington et al., 2015). However, a few issues remain challenging, such as inertia effects from different test methods, specimen dimensions used in different test methods, inconstant strain rates in different individual tests. Numerical simulations can control these effects and have been used extensively (e.g. Li et al., 2009; Du et al., 2018; Huang et al., 2018; Duan et al., 2019b), but an accurate prediction of rock strength depends on a reasonable constitutive model and a reliable model calibration. Therefore, a simple, fast, and data-driven method is needed to predict the compressive strength of rocks at various strain rates and to pro- mote our understanding of the rate-dependent strength of rocks."}], "source_title": "Application of machine learning in predicting the rate-dependent compressive strength of rocks", "source_doi": "10.1016/j.jrmge.2022.01.008", "source_url": "https://www.sciencedirect.com/science/article/pii/S167477552200049X", "source_license": "cc-by-sa", "zotero_item_key": "KNCWQQJD"}
{"id": "ed17f4a99e782a96", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Reinforcement learning based process optimization and strategy development in conventional tunneling; DOI: 10.1016/j.autcon.2021.103701; license: cc-by."}, {"role": "assistant", "content": "Before every move, a number (r) is drawn from a random uniform distribution in the range between 0 and 1. r governs the probability for a move to be a random action (if r ≤ε) or to be based on the agent’s prior experience (if r > ε). Throughout training ε will decay following Eq. (2), where the new ε for the next episode (εi+1) is computed by the previous ε (εi) times a constant – the epsilon decay (εd). εi+1 = εi*εd (2) This process guarantees that the agent can explore the environment in the beginning and shifts towards more exploitation of its knowledge towards the end of training. We set the minimum ε to 0.05 below which there is no more decay and ε will be kept constant (see “final explora­ tion” in Table 2). 3.3. Environment As the DQN agent is the RL pendant to the real life geotechnician, the environment is the RL representation of the rockmass and of logistical processes of the construction site (state) as well as the feedback that is received for better or worse excavation performance (reward). Based on the agent’s actions, its main tasks are to update and yield the current state of the construction-site and to provide feedback. 3.3.1. States The state that is observed by the DQN agent (see Section 3.2) is a hypermatrix of the shape 2 × 2100 × 2 which represents a geological section and a section that shows where support is already installed. The hypermatrix is structured in the following way (see Fig. 6): 2 rows for top heading and bench respectively; 2100 columns for the total length of the tunnel in decimeters plus an additional area beyond the break­ through (see below); 2 channels for the geological section and the sec­ tion with the installed support respectively. The values of the hypermatrix – originally ranging from 0 to 2 (see Section 3.1) - are then scaled between 0 and 1. In the channel of the geological section, 0 therefore translates to “not yet excavated area” and 0.5 and 1 represent gt1 and gt2 respectively (see Table 1 and Section 3.1). In the channel of the already installed support, 0 means “no installed support” and 1 means “installed support”. There are 2 terminal states in the simulation (i.e. states that lead to abortion of the simulation after their occurrence): ▪ A breakthrough is achieved if and only if both posth and posbi are ≥tl. In this case a breakthrough reward is given (see next section) and the simulation is finished. ▪The second terminal state is a timeout which is set to 200 ac­ tions; i.e. the agent must achieve a breakthrough within less than 200 actions or otherwise the simulation is aborted, and a negative breakthrough reward is given. The environment is designed in a way that once posth or posbi are ≥tl, the position of this part of the excavation is not updated anymore even if it is further excavated. We introduced the “timeout” as we observed in the experimental phase of the study that the agent sometimes reaches a breakthrough with the top heading or the bench but keeps on excavating in the already excavated part of the tunnel. On the one hand this causes the episode to be infinitely long, and on the other hand this leads to excessively large negative penalties which negatively affect the training process. We set the total length of the sections to 210 m (i.e. tunnellength + maximum possible length of face support) to give the agent the freedom to use actions that install face support even for the last few blasts. From a “real world tunneling” perspective this does not make sense and in the simulation, this would lead to additional penalties, however, the goal is that the agent learns things like this and is not forced to do so. 3.3.2. Reward system Rewards have the purpose to tell the agent if its actions are beneficial for the total reward. The ultimate goal of the agent is to maximize the return (i.e. sum of the rewards throughout the episode – here: excavate the whole tunnel). During an episode, a reward (also called penalty in case of a negative reward) is given for each individual move that the agent takes [26]. The action is chosen to maximize the expected return of discounted rewards. Reward systems in RL can vary greatly depending on the given task."}], "source_title": "Reinforcement learning based process optimization and strategy development in conventional tunneling", "source_doi": "10.1016/j.autcon.2021.103701", "source_url": "https://www.sciencedirect.com/science/article/pii/S0926580521001527", "source_license": "cc-by", "zotero_item_key": "RXHUVV3X"}
{"id": "ab9b96f25edfda4e", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Automated Excavator Based on Reinforcement Learning and Multibody System Dynamics; DOI: 10.1109/access.2020.3040246; license: cc-by-sa."}, {"role": "assistant", "content": "For more information, see https://creativecommons.org/licenses/by/4.0/ VOLUME 8, 2020\n[page 2] I. Kurinov et al.: Automated Excavator Based on RL and Multibody System Dynamics In view of the complexity of the task and environment, utilization of machine learning can be considered a com- pelling option [3]. Reinforcement learning (RL) is a machine learning method that can be adopted for efficient learning of interactions with a complicated environment. RL uses the concept of an agent – a program capable of performing actions in the environment [4] – to determine actions with the aim of maximizing a pre-defined expected cumulative future reward. The learning process in RL consists of constant exploration of possible actions, acquiring information about the environment from observations, and receiving an evalua- tion of an action by a reward [4]. The observation and reward functions allow programming of the behavior of the model to adapt to changes in the environment: a positive reward is provided for the right behavior and a negative reward given after incorrect behavior. In addition, the agent can be eas- ily retrained for different behavior patterns, which provides opportunities for enhancement of the automated machine. Machine automation can be addressed in multiple ways, as reported in previous studies. In [5], the automation was based on training a fuzzy logic algorithm to mimic the behavior of a skilled human operator. Another study [6] utilized laser rangefinders mounted onto the excavator for automation of digging and loading procedures. In [7], a tele-operated assistant system controlled a real excavator via a wireless local area network. A multibody simulation and behavior-based excavation control approach for application with excavators performing landscaping tasks was introduced in [8]. In recent years, numerous publications have focused on the topic of artificial intelligence in automation. An example from the area of automation of heavy machinery is a study examining the application of Proximal Policy Optimization to teach an agent to operate an excavator [9]. The main goal of the study was to train the neural network to accomplish a leveling task known as bucket dragging. The research used a Dynasty simulation engine (propriety software of Cater- pillar Inc.) in combination with the Open AI Gym frame- work [9]. Another study applied a machine learning approach to automation of a wheel loader [10]. Experimental data were collected on an expert driver filling the bucket of a Volvo L110G wheel loader and a regression model was then optimized with a reinforcement learning algorithm in terms of bucket filling and fuel efficiency [10]. It should be noted that utilization of reinforcement learning for heavy machinery is not limited to controlling the movements of the machine. For example, reinforcement learning has been used in develop- ment of a real-time energy management system of a hybrid excavator to control energy flow and optimize efficiency [11]. The approach of mimicking the behavior of a human oper- ator can have negative effects on efficiency, because the oper- ator might not operate the machine optimally. Methods based on training with a virtual simulation model use simplistic representations of the environment and machine, which may result in difficulties when adopted for implementation on real machines. There is increasing evidence that neural network con- trol policies trained in simulation can also work in the real world if the training scheme is designed correctly [12]–[14]. However, conducting real-world tests on heavy machinery such as an excavator is expensive, and simulation experi- ments are first needed to lay the groundwork. This article extends the only prior study on neural network excavator control [9] by learning to control a more detailed multibody excavator model in a more complex task, demonstrating the scalability of the approach. We also describe the multibody simulation in detail, whereas [9] treats the simulator as a black box. A multibody model is a computer-based simu- lation model replicating the physical structure and behav- ior of a machine with interconnected rigid or flexible bod- ies. Use of a multibody machine model is beneficial for training, because it eliminates risks of damaging the real machine and objects on the construction site and endangering personnel. This study applies five methods for achieving train- ing of the autonomous agent."}], "source_title": "Automated Excavator Based on Reinforcement Learning and Multibody System Dynamics", "source_doi": "10.1109/access.2020.3040246", "source_url": "https://ieeexplore.ieee.org/ielx7/6287639/6514899/09268069.pdf", "source_license": "cc-by-sa", "zotero_item_key": "7F2IQNFD"}
{"id": "b9f4c46b759de121", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Wheel Loader Scooping Controller Using Deep Reinforcement Learning; DOI: 10.1109/access.2021.3056625; license: cc-by-sa."}, {"role": "assistant", "content": "The DDPG network structure. D. DDPG ARCHITECTURE Our implementation of DDPG includes two neural networks (actor and critic), as shown in Fig. 6. Both networks have the current state as input; i.e., 12-dimensional sensor data. VOLUME 9, 2021 24149\n[page 6] O. Azulay, A. Shapiro: Wheel Loader Scooping Controller Using Deep Reinforcement Learning FIGURE 7. Sequence of scooping motions of the Komodo robot with an arm mechanism in simulated and real-life experiments. The actor network output is composed of 3-dimensional deterministic actuator commands that drive the tilt, lift, and the robot’s velocity. The output of the actor network is merged with the first layer of output from the critic network, which in turn is used to output the action-value function Qπ (s, a). The action-value function is used to produce temporal- difference errors deriving the learning in both the actor and critic. We used three hidden layers in both the actor and critic networks. Each layer is fully connected that includes layer normalization whose dimensions decrease from 100, 80, and 60 respectively. A hyperbolic tangent function was used as the activation function to constrain the range of the actions. E. REWARD SHAPING To guide the robot to excavate soil during the task, while not obligating the robot to follow a specific trajectory, we defined a shaped reward: Pt = \u001a 1 −tanh2 (∥dend∥) penetrate 1 −tanh2 (∥dt∥) ¬penetrate \u001b (5) Up to the penetration into the earth; i.e., ¬penetrate, the reward increases while the agent minimizes the relative distance dt, thus encouraging the robot to engage with the pile. During penetration; i.e., penetrate, the reward increases while the robot minimizes the distance to the terminal state, dend. To encourage the robot to scoop more soil during the process, we added a positive reward consisting of mul- tiplying the excavated earth indicator, wt, by a constant cw. Hence, we defined the immediate reward as: rt = \u001a Pt + cwwt ∥dend∥≤ϵ Pt ∥dend∥> ϵ \u001b (6) where ϵ is the distance tolerance to the terminal state. The precondition for penetration into the soil; i.e., penetrate, is defined as: penetrate = minpi dB −dpi ≤1 (7) The values dB, dpi are the robot’s bucket tip and particle i positions, respectively. F. POLICY TRAINING In each episode during the simulation training session, the pile was initialized in a specific location and particle arrangement. Throughout the trial, the robot attempted to complete the bucket-filling process within 8 seconds using combination of the three DoF. We considered the outcome of an episode a success and terminated if, within the maximum episode time, the tip of the loader bucket reached the termi- nal located outside the pile. We trained the networks with algorithmic parameters using conservative values to achieve stable and reliable convergence. The model was trained from random initial weight values using a single Nvidia Quadro P2000D GPU for several trials of 800 episodes which took approximately three hours each. We employed a simple exploration strategy as described in [16] that includes noise inputted to the actor’s action. FIGURE 8. Normalized learning curve for the DDPG agent over 800 episodes (A). The fill factor is represented by the excavated soil in the bucket normalized by full bucket capacity (B). Cycle time represents the operation cycle time from the beginning of the episode until it reaches the terminal state (C). VI. SIMULATIONS AND EXPERIMENTAL RESULTS A. SIMULATION EVALUATION We assessed the learned scooping cycles in the simulation using both the average episode reward, and standard earth- moving parameters; namely, the fill factor and operation cycle 24150 VOLUME 9, 2021\n[page 7] O. Azulay, A. Shapiro: Wheel Loader Scooping Controller Using Deep Reinforcement Learning FIGURE 9. Evaluation of learned scooping cycles in the simulations. Bucket tip trajectory (A), robot velocity (B), lift angle (C) and tilt angle (D). time, as shown in Fig. 8. The scooping policy converged at the end of the training phase with an average loading cycle time of 5[sec] and a fill factor of 65% from full bucket capacity with an overfill (approximately 23 particles). The learned scooping policy can be classified into 3 phases, as indicated in Fig. 9. In the first phase, the robot moves toward the pile of particles while choosing the height and penetration angle to approximately zero tilt and minimum lift angles."}], "source_title": "Wheel Loader Scooping Controller Using Deep Reinforcement Learning", "source_doi": "10.1109/access.2021.3056625", "source_url": "https://ieeexplore.ieee.org/ielx7/6287639/9312710/09344588.pdf", "source_license": "cc-by-sa", "zotero_item_key": "IDZBIHT7"}
{"id": "1872410f2da5eab3", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Automated Excavator Based on Reinforcement Learning and Multibody System Dynamics; DOI: 10.1109/access.2020.3040246; license: cc-by-sa."}, {"role": "assistant", "content": "MULTIBODY SYSTEM DYNAMICS Semi-recursive formulation is an efficient multibody- based method for simulating complex mechanical sys- tems [16]–[19]. It enables dynamic responses to be solved in real-time and sometimes even faster than real-time [16]. Next, the formulation is briefly introduced. Consider two rigid bodies n and n−1 with reference frames located at their center of mass as depicted in Fig.1. The bodies produce an open-chain system with Nb = 2 bodies connected by a joint. The locations of the joint in body n −1 and n are denoted Q and P respectively. The relative displacement FIGURE 1. Multibody system of two rigid bodies. vector between points Q and P can be described using vector dn−1,n. The position of the joint fixture point P in the body n can be expressed as [16]: rn = Rcm n−1 + An−1 ¯un−1 + dn−1,n (1) where Rcm n−1 is the position vector that describes the center of mass of body n−1, An−1 is the rotation matrix of body n−1, and ¯un−1 is the location of point Q with respect to the body frame of reference n−1. The rotation matrix An of body n can be expressed using a concept of successive rotations as [16]: An = An−1An−1,n (2) where An−1,n is the relative rotation matrix between two neighboring bodies. In this study, the rotation of the bodies is expressed by Euler parameters. Velocity of point P can be expressed as [16]: ˙rn = ˙Rcm n−1 + ωn−1 × An−1 ¯un−1 + ˙dn−1,n (3) where ˙Rcm n−1 is the velocity vector of the center mass of body n−1 and ωn−1 is the angular velocity of body n−1. Angular velocity of body n can be expressed as [16]: ωn = ωn−1 + ωn−1,n (4) where ωn−1,n is the relative angular velocity between body n −1 and body n. By employing the principle of virtual work, equations of motion for a multibody system can be expressed [16]: δW = δ ˙qT (M¨q + C −Q) (5) where δW is the virtual work of multibody system, δ ˙q is the vector of 6Nb generalized virtual velocities, M is the 6Nb × 6Nb mass matrix, ¨q is the vector of acceleration of generalized coordinates, C is the quadratic velocity vector, and Q is the vector of generalized external forces. Vectors ¨q, C and Q have 6Nb dimensions and have the form [16]: ¨q = \u0002 ¨qT 1 ¨qT 2 . . . ¨qT Nb \u0003T (6) C = h CT 1 CT 2 . . . CT Nb i (7) 214000 VOLUME 8, 2020\n[page 4] I. Kurinov et al.: Automated Excavator Based on RL and Multibody System Dynamics Q = h QT 1 QT 2 . . . QT Nb i (8) where ¨qn= h ¨RcmT n ˙ωT n iT . The vector of virtual velocity δ ˙q can be related to virtual joint velocities δ˙z as [16]: δ ˙q = Rδ˙z (9) where R is the velocity transformation matrix. For open-loop systems, the joint coordinates represents a set of minimal coordinates. By taking the derivative of (9), generalized accelerations can be expressed as [16]: ¨q = R¨z + ˙R˙z (10) Using (10) and (11), the equations of motion can be rewritten as [16]: δ˙zT RT \u0002 M \u0000R¨z + ˙R˙z \u0001 + C −Q \u0003 = 0 (11) (11) also holds for independent virtual velocities δ˙z; there- fore, virtual velocities can be eliminated [16]: RT MR¨z = RT (Q −C) −RT M ˙R˙z (12) The velocity transformation matrix R plays an important role in the formulation as it directly affects the efficiency of the method. One of the most efficient ways to compute the velocity transformation matrix R is application of the element-by-element technique [17]: RNb = [R1 Nb R2 Nb . . . RPb Nb] (13) Closed-loop systems can be computed by expressing the system in the form of an open-loop system and then closing it be employing constraints [16]. Accordingly, the equation of motion takes the form [17]: (RT MR + αφT z φz)¨z = RT (Q −C) −RT M ˙R˙z −αφT z \u0010 β2φ + µ ˙φ + ˙φz˙z + ˙φt \u0011 (14) where α is a constant penalty factor, β2 and µ are constants representing natural frequency and damping ratio respec- tively, φ is the loop closure constraint equations, φz is a Jacobian matrix of the constraints, and ˙φ and ¨φ are the first and second order derivatives of the constraint equations. C. HYDRAULICS In this study, the hydraulic system is modelled using the lumped fluid theory [20]. The approach is based on partition- ing the system into discrete volumes in which the pressure is assumed to be equally distributed. The pressure in one volume can be computed as: ˙ps = Bes Vs Xnc k=1 Qsk (15) where Vs is a volume, Qsk is the sum of the flows going in or out of the volume, nc is the number of hydraulic compo- nents belonging to the volume, and Bes is the effective bulk modulus corresponding to the volume."}], "source_title": "Automated Excavator Based on Reinforcement Learning and Multibody System Dynamics", "source_doi": "10.1109/access.2020.3040246", "source_url": "https://ieeexplore.ieee.org/ielx7/6287639/6514899/09268069.pdf", "source_license": "cc-by-sa", "zotero_item_key": "7F2IQNFD"}
{"id": "efce9b6d3881d481", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Cutterhead and Cutting Tools Configurations in Coarse Grain Soils; DOI: 10.2174/1874836801711010182; license: cc-by."}, {"role": "assistant", "content": "[page 1] See discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/316320965 Cutterhead and Cutting Tools Configurations in Coarse Grain Soils Article in The Open Construction and Building Technology Journal · April 2017 DOI: 10.2174/1874836801711010182 CITATIONS 4 READS 2,310 3 authors, including: Xinggao li Beijing Jiaotong University 24 PUBLICATIONS 368 CITATIONS SEE PROFILE All content following this page was uploaded by Xinggao li on 10 August 2017. The user has requested enhancement of the downloaded file.\n[page 2] Send Orders for Reprints to reprints@benthamscience.ae 182 The Open Construction and Building Technology Journal, 2017, 11, 182-199 1874-8368/17 2017 Bentham Open The Open Construction and Building Technology Journal Content list available at: www.benthamopen.com/TOBCTJ/ DOI: 10.2174/1874836801711010182 RESEARCH ARTICLE Cutterhead and Cutting Tools Configurations in Coarse Grain Soils Xinggao Li 1,*, Dajun Yuan 1 and Qingfei Huang 2 1 School of Civil Engineering, Beijing Jiaotong University, Beijing 100044, China 2 CCCC Highway Consultants CO., Ltd, Beijing 100088, China Received: November 30, 2016 Revised: February 13, 2017 Accepted: February 13, 2017 Abstract: Background: Many subway/metro tunnels were constructed using close face shield machines in coarse grain soils in last decades. Configurations of the shield cutterhead and cutting tools are critical in terms of cost and project duration. There are currently no recognized methodologies available for the configurations due to the various geological and technical influencing factors. Several aspects have been considered in many single case studies. Objective and Method: To aid a better understanding for the involved design and operational aspects in a system way, main physical and mechanical characteristics of the coarse grain soils encountered in China are briefly summarized; advantages and disadvantages of the spoke type and plate type cutterhead equipped with corresponding cutting tools are outlined; performances of the different shield cutterheads used in the typical subway construction cases are introduced. Results: Based on the jobsite experiences, discussed are the options with respect to design of the cutterhead and cutting tools in the coarse grain soils, including opening ratio of the cutterhead, configuration of the cutting tools, use of the single disc cutters with inserts, reasonable movement of shield machine, robust shield machines against wears and interventions for changing tools. Conclusion: According to the discussions, the rule-of-thumb suggestions are presented to provide references for later similar projects. Keywords: Cutterhead, Cutting tools, Coarse grain soils, Opening ratio, Disc cutter, Wear. 1. INTRODUCTION In the largest cities of China, the subway/metro systems have been increased dramatically due to the large population and limited surface spaces in recent years. So far, most of the subway tunnels in the cities have been constructed using the Earth Pressure Balanced (EPB) shield tunneling methods because of the many advantages that they offer over the conventional methods. Due to alignment constraints, many subway tunnels in the cities such as Beijing, Shenyang and Chengdu, are buried in whole or part in the coarse grain ground mainly consisting of fine/medium coarse sand and sandy gravel/cobble stratum with boulders. Various difficulties are encountered in the shield tunneling work undertaken in such ground conditions. Particularly, when tunneling in the cobble and boulder laden ground, the cutting tools suffer excessive wears and tears, and are liable to damage, fail and pull out. And even worse the cutterhead is worn through, which will adversely affect the rapid construction of tunnels. Consequently, it is of great significance to study the reasonable configurations of the cutterhead and cutting tools that is appropriate for the * Address correspondence to this author at the School of Civil Engineering, Beijing Jiaotong University, No.3 of Shangyuancun, Xizhimenwai, Haidian District, Beijing, China; Tel: (86-10) 51688115; Fax: (86-10) 51688022; E-mails: lixg@bjtu.edu.cn, lxg_njtu@163.com\n[page 3] Cutterhead and Cutting Tools Configurations The Open Construction and Building Technology Journal, 2017, Volume 11 183 shield tunneling in coarse grain soils."}], "source_title": "Cutterhead and Cutting Tools Configurations in Coarse Grain Soils", "source_doi": "10.2174/1874836801711010182", "source_url": "https://openconstructionbuildingtechnologyjournal.com/VOLUME/11/PAGE/182/PDF/", "source_license": "cc-by", "zotero_item_key": "7KDIA2X2"}
{"id": "27faf5371857993d", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Reinforcement learning based process optimization and strategy development in conventional tunneling; DOI: 10.1016/j.autcon.2021.103701; license: cc-by."}, {"role": "assistant", "content": "A copy of the agent is saved after every 1000 episodes. To observe how big the differences are between individual training runs, we trained several identical DQN agents in the above described environment. After every episode, 21 parameters are saved to monitor the training progress. Below, the parameters that are mentioned in the paper are given (see the code “A_utilities.py” in Appendix 1 for a list of all recorded parameters): ▪ number of the current episode ▪ cumulative reward of the episode ▪ current value of ε ▪ the number of face instabilities of the whole episode ▪ average loss of the DQN agent throughout the episode ▪ number of moves/blasts that were required to finish the episode/to reach a terminal state ▪ 8 counters for how many times each of the actions of Table 3 were used; the goal of these counters is to see if the agent favors some actions over others and to detect “strategy changes” 5. Experiments In this section we present five training paths of exemplary agents and discuss the different strategies they found to deal with the given task. In Fig. 7, recordings of the agents’ training are given which shows that each agent has found a unique solution, and all training paths are substan­ tially different from one another. Comparing the five agents to one another, all of them were able to increase the cumulative reward per episode above 200 within 10,000 episodes (Fig. 7 first row). After around 30,000 episodes, first differ­ ences arise where first the cumulative reward of agent 2 and then of agent 3 started to stagnate. The reward of agent 1 stagnates at around 80,000 episodes. The cumulative rewards of the agents 4 and 5 kept on increasing, whereas agent 5 became instable after around 75,000 epi­ sodes and agent 4 reached the maximum reward at the end of the 120,000 training episodes. Right from the start, all agents started to use long advance lengths as a means to decrease the number of blasts/moves per episode as this is an effective way to maximize the achievable reward (Fig. 7 second row). Where the agents 1, 4 and 5 all took a similar strategy that aims at continuously minimizing the blasts/moves per episode throughout the whole training, agents 2 and 3 reached a minimum of around 120 blasts/ moves per episode after around 30,000 episodes. This correlates well with the stagnating rewards after 30,000 episodes of agents 2 and 3 as described above. The biggest differences in training paths can be observed with Table 4 Rewards that the agent receives from the environment in response to its actions. Reward (points) Description Condition tl * 3 reward for achieving breakthrough posth ≥tl and posbi ≥tl tl * 3 * -1 penalty for a timeout if number of moves in current episode is >200 −6 penalty for using the wrong excavation sequence, i.e. the bench is driven further ahead than the top heading posbi > posth −5 penalty for unstable tunnel face conditions determined from face pressure pf as evaluated by Eq. (3) (see below) pf ≥0 −4 penalty for changing from top heading to bench excavation or vice versa, as this usually involves a delay of the excavation due to logistics if current a ≥200 and prev. a < 200 or if current a < 200 and prev. a ≥200 −3 penalty for a too far distance between top heading and bench (see Section 3.1) if posth – posbi > distmax (initially set to 50 m) −2 penalty for using face support as this consumes additional time and resources if a = 112; a = 152; a = 202; a = 222 −1 penalty for every other move that does not meet any of the above conditions no other condition is fulfilled G.H. Erharter et al.\n[page 8] Automation in Construction 127 (2021) 103701 8 respect to face instabilities, where only the agents 2 and 3 started to actively decrease the number of face instabilities (Fig. 7 third row) early in the training process (within 10,000 episodes). In contrast to that, the number of face instabilities increased within the first 10,000 episodes for the other agents and then only decreased slowly throughout the rest of the training. The best performing agent 4 shows a remarkable trend of stagnating face instabilities until around 80,000 episodes, followed by a decrease towards the end of training. While the goal was to let all agents train for 120,000 episodes, the training process of the agents 2 and 5 became unstable after around 75,000 and 85,000 episodes respectively, which led to the abortion of training after it could be observed that the agent would not recover from this."}], "source_title": "Reinforcement learning based process optimization and strategy development in conventional tunneling", "source_doi": "10.1016/j.autcon.2021.103701", "source_url": "https://www.sciencedirect.com/science/article/pii/S0926580521001527", "source_license": "cc-by", "zotero_item_key": "RXHUVV3X"}
{"id": "80acc50861df3d3e", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Two-phase multi-model automatic brain tumour diagnosis system from magnetic resonance images using convolutional neural networks; DOI: 10.1186/s13640-018-0332-4; license: cc-by-sa."}, {"role": "assistant", "content": "Deep learning methods provide high efficiency and the ability to process numerous MRIs from databases [35]. This study focuses on CNNs, which have gained popularity among researchers for object recognition and biological image segmentation. The developed diagnosis system has two phases for tumour detection and localization. The first phase con- verts tumour detection into an MRI classification problem and then categorizes MRIs into normal and abnormal images. The second phase focuses on tumour localiza- tion inside the abnormal images. Figure 1 shows a block diagram of the proposed system. MRIs are delivered to the system’s input, and then, the preprocessing tasks are performed. A feature extraction tool is used to select the relevant features, and a classification tool is applied to decide whether these features belong to a normal or an abnormal MRI. To achieve accurate results, the five- layer R-CNN has been trained and evaluated using MRIs from BraTS 2013 as a standard benchmark for tumour localization. 3.1 Brain tumour detection phase 3.1.1 Input image and preprocessing Due to the ability of magnetic resonance imaging to provide information about the soft tissues in the brain, databases of MRIs were used in this study as an input to the developed system. The goal of the preprocessing stage is to ensure the MRI readiness for the next stage. It improves image quality, reduces the noise level, and enhances the overall image resolution by employing clas- sical filtering techniques [36]. The preprocessing stage also adjusts the image size and type to a 227 × 227 RGB image, which is provided to the trainable CNN in the feature extraction stage.\n[page 4] Abd-Ellah et al. EURASIP Journal on Image and Video Processing (2018) 2018:97 Page 4 of 10 Fig. 1 A flowchart of the developed CNN-based system for brain tumour detection and localization. The flowchart shows the two system phases (detection and localization phases) and highlights the two different types of convolutional neural network (CNN) used for each system phase. The AlexNet, VGG-16, and VGG-19 models fit the CNN in the detection phase, and a five-layer fully developed R-CNN was used in the localization phase 3.1.2 Convolutional neural network (CNN) CNN is the deep learning model that has improved over the past two decades. CNNs can learn features automati- cally from input data, especially images, as is the case in this research [37]. The convolutional layers are used to convolve the input image with kernels (weights) to obtain a feature map. The weights of the kernels connect the fea- ture map units to the previous layer. AlexNet and two other CNN models, namely, VGG-16 and VGG-19, were used and evaluated in the tumour detection phase. How- ever, the focus of the theoretical foundation section is on the AlexNet model. The AlexNet model was selected due to its flexibility to be modified, its ability to reduce over- fitting using a dropout layer, and its capability to train faster through using a rectified linear unit (ReLU). The AlexNet model utilized for tumour detection is a pre-trained network from the MatConvNet toolbox [12]. It consists of 23 layers with weights. The first layer defines the input dimensions, including an input image size of 227 × 227 × 3. A series of convolutional layers intervene with ReLU and max-pooling layers. The final layer is a classification layer with 1000 classes. The tumour detection phase employs the CNN network for feature extraction, and there are few layers in the CNN that are suitable for this goal. The CNN depends on nine layers, as shown in Fig. 2: an image input, two convolution, two ReLU, two cross channel normalization (Norm), and two max-pooling layers. Layers 2 and 6 are convolutional layers with 11×11×3 and 5×5×48 convolutions, respec- tively. If we suppose that the input x and the weight w are realized on integer t, the 2D discrete convolution can be defined as in Eq. 1: y(i, j) = x(i, j) ∗w(i, j) = \u0002 m \u0002 n x(m, n)w(i −m, j −n) (1) where m ∗n is the size of the convolution matrix, w and x are the inputs from the previous layer, and y is the output of the convolutional layer. Table 2 The configurations of the MRIs database in the training and testing datasets used in tumour detection phase Total images Distribution in the training set Distribution in the testing set Normal Abnormal Normal Abnormal 349 45 77 64 163\n[page 5] Abd-Ellah et al. EURASIP Journal on Image and Video Processing (2018) 2018:97 Page 5 of 10 Fig."}], "source_title": "Two-phase multi-model automatic brain tumour diagnosis system from magnetic resonance images using convolutional neural networks", "source_doi": "10.1186/s13640-018-0332-4", "source_url": "https://jivp-eurasipjournals.springeropen.com/track/pdf/10.1186/s13640-018-0332-4", "source_license": "cc-by-sa", "zotero_item_key": "2ZZSU54R"}
{"id": "52e344580d4946e0", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Reinforcement learning based process optimization and strategy development in conventional tunneling; DOI: 10.1016/j.autcon.2021.103701; license: cc-by."}, {"role": "assistant", "content": "Chollet (Ed.), Deep Learning with Python, Manning, Shelter Island, NY, 2018. [44] M. Abadi, A. Agarwal, P. Barham, E. Brevdo, Z. Chen, C. Citro, G.S. Corrado, A. Davis, J. Dean, M. Devin, S. Ghemawat, I. Goodfellow, A. Harp, G. Irving, M. Isard, Y. Jia, R. Jozefowicz, L. Kaiser, M. Kudlur, J. Levenberg, D. Man´e, R. Monga, S. Moore, D. Murray, C. Olah, M. Schuster, J. Shlens, B. Steiner, I. Sutskever, K. Talwar, P. Tucker, V. Vanhoucke, V. Vasudevan, F. Vi´egas, O. Vinyals, P. Warden, M. Wattenberg, M. Wicke, Y. Yu, X. Zheng, TensorFlow: Large-Scale Machine Learning on Heterogeneous Systems. https://www.tensor flow.org/, 2015. [45] F. Chollet, Others, Keras, 2015. [46] ¨Osterreichisches Normungsinstitut, Untertagebauarbeiten: Teil 1: Zyklischer Vortrieb, Wien 91.010.20; 93.020, 2019. [47] Z. Zhao, Y. Liang, X. Jin, Handling large-scale action space in deep Q network, in: 2018 International Conference on Artificial Intelligence and Big Data (ICAIBD), IEEE, Chengdu, 2018, pp. 93–96. [48] T. Zahavy, M. Haroush, N. Merlis, D.J. Mankowitz, S. Mannor, Learn what not to learn: action elimination with deep reinforcement learning, in: Advances in Neural Information Processing Systems 31 (NIPS 2018), Montreal, Canada, 2018, pp. 3566–3577. [49] ¨OGG, Guideline for the Geotechnical Design of Underground Structures with Conventional Excavation: Ground characterization and coherent procedure for the determination of excavation and support during design and construction. Translated from version 2.1, second.first, Salzburg, 2010. [50] H. Lauffer, Gebirgsklassifizierung für den Stollenbau, Geol. Bauwesen 24 (1958) 46–51. [51] P. Zhang, H. Li, Q.P. Ha, Z.-Y. Yin, R.-P. Chen, Reinforcement learning based optimizer for improvement of predicting tunneling-induced ground responses, Adv. Eng. Inform. 45 (2020) 101097, https://doi.org/10.1016/j.aei.2020.101097. [52] J. van Eldert, H. Schunnesson, D. Johansson, D. Saiang, Application of measurement while drilling technology to predict rock mass quality and rock support for Tunnelling, Rock Mech. Rock. Eng. 53 (2020) 1349–1358, https://doi. org/10.1007/s00603-019-01979-2. [53] A. Radinger, F. Fasching, G. Pack, I. Kreutzer, D. Kostial, Consistent exploration by probe drilling and TSWD through the example of the Koralm tunnel/Konsequente Vorauserkundung mittels Bohrungen und TSWD am Beispiel des Koralmtunnels, Geomechanik Tunnelbau 7 (2014) 540–550, https://doi.org/10.1002/ geot.201400038. G.H. Erharter et al."}], "source_title": "Reinforcement learning based process optimization and strategy development in conventional tunneling", "source_doi": "10.1016/j.autcon.2021.103701", "source_url": "https://www.sciencedirect.com/science/article/pii/S0926580521001527", "source_license": "cc-by", "zotero_item_key": "RXHUVV3X"}
{"id": "d60649ea8a0dc854", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Reinforcement learning based process optimization and strategy development in conventional tunneling; DOI: 10.1016/j.autcon.2021.103701; license: cc-by."}, {"role": "assistant", "content": "In this paper we present a novel RL based framework for construction process optimization and strategy development for conventional * Corresponding author. E-mail address: erharter@tugraz.at (G.H. Erharter). Contents lists available at ScienceDirect Automation in Construction journal homepage: www.elsevier.com/locate/autcon https://doi.org/10.1016/j.autcon.2021.103701 Received 1 December 2020; Received in revised form 29 March 2021; Accepted 2 April 2021\n[page 2] Automation in Construction 127 (2021) 103701 2 tunneling. Such models can act as decision support for the geotechnical engineer, engineering geologist, geotechnician etc. (hereafter “geo­ technician” is used) (design choices, progress-planning) and in the long run such models work towards full automation in underground con­ struction. Hence, the model is a first attempt to automate decisions made by the geotechnician on face in underground construction. In the next section (2) we frame the process of conventional tunneling as a RL problem and provide details on how these two disci­ plines can connect. Section 3 is the main methodological section that presents the geotechnical scenario at the background of this RL simu­ lation as well as the agent and the environment. In Section 4 we describe the training process and in Section 5 we show experiences gathered during the training and testing of agents. A conclusion and implications for the vision of “digital tunneling” is given in Section 6 and we present an outlook in the last Section 7. A reference to the Python based code for this paper, is given in the appendix. 2. Conventional tunneling as a reinforcement learning problem The process of reinforcement learning (RL) is typically depicted as a closed loop where an agent takes different actions, to influence an environment which responds by sending an updated state as well as a reward signal to the agent (e.g. [26]). To apply RL, the learning problem must undergo the Markov property, i.e. we only need to know the cur­ rent state of the system, to make a decision [27]. Therefore, the state must include information about all aspects of the past agent–environ­ ment interaction that make a difference for the future. This can be said to be true for tunnel excavation where we only need to know the state of the rockmass and the excavation-process to decide on how to proceed. Tunnel construction follows several cyclic and sequential processes, some of which can be framed as loops and therefore translated to RL problems. The most outstanding loop in conventional tunneling is the excavation of an underground opening with a sequential construction process of: blasting, mucking and rock support installation [28] (“excavation loop” in Fig. 1). While the components of this excavation loop are themselves often sequential processes (e.g. blasting sequence, support installation etc.), the excavation loop specifically is based on a sequence of geotechnical decisions belonging to a bigger cycle which we refer to as the excavation sequence decision – loop. Looking at the bigger picture, excavation sequence decisions are one part of the whole con­ struction phase of a tunnel and therefore part of the whole tunnel life cycle (Fig. 1). The focus of this study is to create a simulation of a simplified version of the process that governs the general excavation sequence decisions and frame it as a RL-loop. This process can be translated to a RL-loop/ Markov decision process [27] consisting of the following components: ▪ the decision making geotechnician is the agent ▪ processes like “top heading excavation”, “bench excavation”, “installation of face support” etc. are the actions ▪the rockmass itself and the construction site with all its pro­ cesses are the environment (here described at each timestep by the state of the environment and the reward-system) ▪the sum of all delays (planned and unplanned) and complica­ tions throughout the course of the excavation which result from the geotechnician’s actions are the reward ▪ the current state of the excavation including information about the past and recent rockmass conditions as well as the already installed support are the state Fig. 2 is a graphical representation of this process which we refer to as “TunnRL” (Tunnel automation with Reinforcement Learning). The individual components of this loop in the above given list as well as in Fig. 1 and Fig. 2 are only for explanatory purpose and do not claim to be complete."}], "source_title": "Reinforcement learning based process optimization and strategy development in conventional tunneling", "source_doi": "10.1016/j.autcon.2021.103701", "source_url": "https://www.sciencedirect.com/science/article/pii/S0926580521001527", "source_license": "cc-by", "zotero_item_key": "RXHUVV3X"}
{"id": "068586fb83c1c57f", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Cutterhead and Cutting Tools Configurations in Coarse Grain Soils; DOI: 10.2174/1874836801711010182; license: cc-by."}, {"role": "assistant", "content": "In view of the high abrasivity of the sandy cobble/gravel soils with boulders, the primary wear on cutting tools and the secondary wear on the support structures such as the cutterhead, if not controlled, are\n[page 5] Cutterhead and Cutting Tools Configurations The Open Construction and Building Technology Journal, 2017, Volume 11 185 excessive, and especially the cutting teeth made of hard metal (cemented carbide) is liable to flake off when subject to impacts of the over-sized particles. Being loosely cemented and with less cohesion, most coarse grain soils exposed are typically unstable strata and 3. their mechanical properties are unfavorable for force generation and transmission, and the ground surface is liable to cave in if improper manipulations of the shield machines exist. Fig. (1). Typical sandy gravel/cobble soil with boulders encountered in Beijing area. 3. CUTTERHEADS, CUTTING TOOLS CONFIGURATIONS AND SHIELD CUTTERHEAD PERFORMANCES 3.1. Cutterheads and Cutting Tools Configurations Advance of a shield machine is heavily dependent on its cutterhead as well as the combination and arrangement of the installed cutting tools. Generally, as shown in Figs. (2 & 3), there are basically two types of cutterheads for soft ground shield machines; one is the spoke type and the other is the plate type. The most significant difference between the two types of cutterheads is the discrepancy in opening ratio of the cutterheads, which is the percentage of openings of the total face area of a cutterhead. The advantages and disadvantages related to opening ratio of the two type’s cutterheads in excavating tunnels are displayed in Table 1. Fig. (2). Spoke type cutterhead.\n[page 6] 186 The Open Construction and Building Technology Journal, 2017, Volume 11 Li et al. Fig. (3). Plate type cutterhead. Table 1. Advantages and disadvantages of the spoke type and plate type cutterhead. Items Spoke Type Cutterhead Plate Type Cutterhead Pressure control of the cutting face Being almost equal to the earth chamber pressure, the cutting face pressure is easy to be controlled. On account of the indirect transmission between earth chamber pressure and cutting face pressure, it is a little difficult to perform precise pressure control of the cutting face. Entry of muck into the excavation chamber Entry of the excavated materials into cutterhead is smooth on account of the high opening ratio, almost resulting in no clogging of the cutterhead. Affected by the plate, the entry of the excavated materials is not smooth and the cutterhead is liable to be clogged and jammed by the excavated materials. Torque of cutterhead The required torque of the cutterhead is small and the manufacturing price of the equipment is low. Necessitate high torque of the cutterhead and large capacity of the equipment, as leads to a high construction cost. Wear of the cutterhead and cutting tools The high opening ratio of the cutterhead causes the relatively minor wears of the cutterhead and cutting tools. Because of the small opening ratio of cutterhead, the excessive wear of the cutterhead and cutting tools is a hackneyed phenomon. Adaptability to the over- sized particles Over-sized grains or boulders that can’t be extracted from the excavation chamber by the screw conveyor must be handled by manual work. Kept out of the cutterhead, the over-sized grains or boulders can be broken by disc cutters. Safety of changing tools The high opening ratio of the cutterhead to some extent decreases the stability of the cutting face. The plate is helpful to stability of the cutting face, and a higher construction safety results. Requirements for shield machine manipulations The requirements are relatively high. The requirements are relatively low. The prevalent excavation tools used for the spoke type cutterhead include tooth bits (scraper, ripper), shell bits, precut bits (ripper) and a center fish-tail bit (nose cone ripper). The idea of this combination of tools is to loosen ground before cutterhead with the center fish-tail bit and the precut bits, and rip off the over-sized stones out of the soil matrix with rippers, and shave off the weak materials with scrapers."}], "source_title": "Cutterhead and Cutting Tools Configurations in Coarse Grain Soils", "source_doi": "10.2174/1874836801711010182", "source_url": "https://openconstructionbuildingtechnologyjournal.com/VOLUME/11/PAGE/182/PDF/", "source_license": "cc-by", "zotero_item_key": "7KDIA2X2"}
{"id": "016ab9a10ddb9f19", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Cutterhead and Cutting Tools Configurations in Coarse Grain Soils; DOI: 10.2174/1874836801711010182; license: cc-by."}, {"role": "assistant", "content": "Bilgin et al. (2012) performed some theoretical estimations to estimate the effect of replacing disc cutters with chisel tools on performance of a tunneling machine in difficult ground conditions based on field measurements [18]. Willis (2012) showed the advances of quality and configuration for the cutting tools developed by Robbins engineers to reduce cutterhead wear at the EPB shield machines [19]. Havekost et al. (2013) introduced a plan to remove the close face shield machine and replace it with a digger shield when constructing the 1829 m long Portsmouth Force Main Tunnel in Portland and the machine was unable to advance through running soils containing cobbles and boulders [20]. Köppl and Thuro (2013) set up a prognosis model for cutter life in soft ground for mixshield machines based on empirical correlations with soil parameters and machine design and advance parameters after the detailed data analysis of the 18 mixshield drives [21]. Duhme and Edelmann (2014) presented the general principles of inspection and monitoring technology developed by Herrenknecht of the shield machine cutting tools as well as discussed examples of their applications [22]. Roby and Willis (2014) analyzed the record-breaking and high-performing projects to identify factors that contribute to fast advance of the EPB shield machines, such as adequate cutting tools and cutterhead design, and an effort was made to form simple, high level guidelines for the optimal shield machine excavation in mixed ground conditions [23]. In-door tests can only be used to deal with limited interaction features and sometimes are not realistic because of the ignorance\n[page 4] 184 The Open Construction and Building Technology Journal, 2017, Volume 11 Li et al. of some basic factors. Numerical simulation seems to be a promising way to capture the interaction between tunneling shields and soils, but its application is greatly limited in practical engineering on account of difficulty in determining constitutive relations of the coarse grain soils. Field observations remain the commonly recognized approach for cutterhead performance prediction. The soft ground shield machine cutterhead design has historically been driven by the iterative designs based on empirical data, observations over various projects and industry rules of thumb [3]. Realistic performance predication of the EPB shield is only possible on the basis of the specific project data [17]. Any in-situ study conducted in the lately finished shield tunnel projects in coarse grain soils is valuable to promote the work of the cutterhead and cutting tools configurations. This issue of the cutterhead and cutting tools configurations for the tunneling shields in coarse grain soils has caused widespread concern in the last decade in China with the completion of many subway shield tunnels, and a growing attention and interest is focused on it. Methodologies for shield tunneling in coarse grain soils progress steadily and some experiences of the cutterhead selection and cutting tools configurations have been published."}], "source_title": "Cutterhead and Cutting Tools Configurations in Coarse Grain Soils", "source_doi": "10.2174/1874836801711010182", "source_url": "https://openconstructionbuildingtechnologyjournal.com/VOLUME/11/PAGE/182/PDF/", "source_license": "cc-by", "zotero_item_key": "7KDIA2X2"}
{"id": "32cf86d35b291235", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Spectral Angle Mapping and AI Methods Applied in Automatic Identification of Placer Deposit Magnetite Using Multispectral Camera Mounted on UAV; DOI: 10.3390/min12020268; license: cc-by-sa."}, {"role": "assistant", "content": "[page 1] \u0001\u0002\u0003\u0001\u0004\u0005\u0006\u0007\b \u0001\u0002\u0003\u0004\u0005\u0006\u0007 Citation: Sinaice, B.B.; Owada, N.; Ikeda, H.; Toriya, H.; Bagai, Z.; Shemang, E.; Adachi, T.; Kawamura, Y. Spectral Angle Mapping and AI Methods Applied in Automatic Identification of Placer Deposit Magnetite Using Multispectral Camera Mounted on UAV. Minerals 2022, 12, 268. https://doi.org/ 10.3390/min12020268 Academic Editor: Yosoon Choi Received: 21 January 2022 Accepted: 18 February 2022 Published: 20 February 2022 Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affil- iations. Copyright: © 2022 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/). minerals Article Spectral Angle Mapping and AI Methods Applied in Automatic Identification of Placer Deposit Magnetite Using Multispectral Camera Mounted on UAV † Brian Bino Sinaice 1,* , Narihiro Owada 2, Hajime Ikeda 1 , Hisatoshi Toriya 1 , Zibisani Bagai 3, Elisha Shemang 4, Tsuyoshi Adachi 1 and Youhei Kawamura 5 1 Department of Geosciences, Geotechnology and Materials Engineering for Resources, Graduate School of International Resource Sciences, Akita University, Akita 010-8502, Japan; ha2ikeda@gmail.com (H.I.); toriya@gipc.akita-u.ac.jp (H.T.); adachi.t@gipc.akita-u.ac.jp (T.A.) 2 Faculty of International Resource Sciences, Technical Division, Akita University, Akita 010-8502, Japan; owada@gipc.akita-u.ac.jp 3 Department of Geology, University of Botswana, Private Bag UB 0022, Gaborone, Botswana; bagaizb16@gmail.com 4 Department of Earth and Environmental Science, Botswana International University of Science and Technology, Private Bag 16, Palapye, Botswana; shemange@biust.ac.bw 5 Faculty of Engineering, Division of Sustainable Resources Engineering, Hokkaido University, Hokkaido 060-8628, Japan; kawamura@eng.hokudai.ac.jp * Correspondence: bsinaice@rocketmail.com † This article is an expanded version this conference paper: Sinaice, B.B.; Takanohashi, Y.; Owada, N.; Utsuki, S.; Hyongdoo, J.; Bagai, Z.; Shemang, E.; Kawamura, Y. Automatic magnetite identification at Placer deposit using multi-spectral camera mounted on UAV and machine learning. In Proceedings of the 5th International Future Mining Conference 2021—AusIMM 2021, Online, 6–10 December 2021; pp. 33–42; ISBN 978-1-922395-02-3. Abstract: The use of drones in mining environments is one way in which data pertaining to the state of a site in various industries can be remotely collected. This paper proposes a combined system that employs a 6-bands multispectral image capturing camera mounted on an Unmanned Aerial Vehicle (UAV) drone, Spectral Angle Mapping (SAM), as well as Artificial Intelligence (AI). Depth possessing multispectral data were captured at different flight elevations. This was in an attempt to find the best elevation where remote identification of magnetite iron sands via the UAV drone specialized in collecting spectral information at a minimum accuracy of +/−16 nm was possible. Data were analyzed via SAM to deduce the cosine similarity thresholds at each elevation. Using these thresholds, AI algorithms specialized in classifying imagery data were trained and tested to find the best performing model at classifying magnetite iron sand. Considering the post flight logs, the spatial area coverage of 338 m2, a global classification accuracy of 99.7%, as well the per- class precision of 99.4%, the 20 m flight elevation outputs presented the best performance ratios overall. Thus, the positive outputs of this study suggest viability in a variety of mining and mineral engineering practices. Keywords: UAV; remote sensing; hyperspectral imaging; multispectral imaging; spectral angle mapping; artificial intelligence; machine learning; deep learning 1. Introduction Safety, cost efficiency and overall system optimization via artificially intelligent (AI) machine learning (ML) as well as deep learning (DL) algorithms are goals set by modern day industries that invest heavily in output maximization [1]. To improve safety and ore tracking abilities, the mining industry has been no stranger to this trend as the introduction of machines such as automated haul trucks and conveyer belts has been seen [1,2]."}], "source_title": "Spectral Angle Mapping and AI Methods Applied in Automatic Identification of Placer Deposit Magnetite Using Multispectral Camera Mounted on UAV", "source_doi": "10.3390/min12020268", "source_url": "https://www.mdpi.com/2075-163X/12/2/268/pdf?version=1645347690", "source_license": "cc-by-sa", "zotero_item_key": "VPIE7856"}
{"id": "b79533c313c599a4", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Two-phase multi-model automatic brain tumour diagnosis system from magnetic resonance images using convolutional neural networks; DOI: 10.1186/s13640-018-0332-4; license: cc-by-sa."}, {"role": "assistant", "content": "5 Conclusions This paper has proposed a two-phase multi-model deep learning-based system for brain tumour detection and localization from MRIs. The main goals of this study are to classify MRIs into normal and abnormal images accord- ing to the absence or presence of a brain tumour and to accurately localize the tumour within the abnormal MRIs. The first system phase employed CNN and ECOC- SVM approaches for feature extraction and classification, respectively. A five-layer R-CNN was used for tumour localization in the second system phase. The tumour detection phase was evaluated using 349 MRIs extracted from the RIDER Neuro MRI database. The empirical work has shown that the method achieved an accuracy of 99.55%. The tumour localization phase was evaluated using the BraTS 2013 database, and a DICE score of 0.87 has been achieved using 804 3D MRIs. The accom- plished results proved the superiority of the proposed deep learning-based method for tumour detection and demonstrate the superiority of the whole system’s perfor- mance concerning both tumour detection and localization measures. Abbreviations ANN: Artificial neural network; BPN: Back propagation neural network; BraTS: Brain tumour segmentation; CAD: Computer-aided diagnosis; CNN: Convolutional neural network; DNN: Deep neural network; DWTs: Discrete wavelet transforms; ECOC: Error-correcting output codes; ECOC-SVM: Error-correcting output codes support vector machine; FCM: Fuzzy C-means; FCNN: Fully convolutional neural network; FCR-NN: Fully convolutional residual neural network; GLCM: Gray-level co-occurrence matrix; HADA: Hybrid abnormality detection algorithm; HG: High grade; ICA: Independent component analysis; k-NN: K-nearest neighbours; KSVM: Kernel support vector machines; LG: Low grade; MRIs: Magnetic resonance images; NPV: Negative predictive value; PCA: Principal component analysis; PNN: Probabilistic neural network; PPV: Positive predictive value; RBFN: Radial basis function neural network; R-CNN: Region-based convolutional neural network; ReLU: Rectified linear unit; RIDER: Reference Image Database to Evaluate Response; SOMNN: Self organizing map neural network; VGG: Visual Geometry Group. Acknowledgements The authors are very thankful to the editor and the anonymous reviewers for their valuable comments and suggestions. Availability of data and materials Not applicable. The magnetic resonance images used in this study are public databases that are cited within the text. Authors’ contributions MA was the main contributor to the ANNs design, implementation, and evaluation. He worked on the MRI databases preparation and the manuscript writing as well. AA focused on the technical results and developed comparisons with the state-of-the-art. He also added valuable efforts into preparing the manuscript. AK worked on the mathematical models of ANNs and offered help in the empirical work done. He also provided a strong support to the manuscript writing process. HH supervised the research outcomes and gave help in the manuscript writing and editing processes. HH provided feedback on the whole structure of the manuscript as well. All authors read and approved the final manuscript. Ethics approval and consent to participate Not applicable. Competing interests The authors declare that they have no competing interests. Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Author details 1Electronic and Communication Department, Al-Madina Higher Institute for Engineering and Technology, Giza, Egypt. 2Department of Computer Science, Electrical and Space Engineering, Luleå University of Technology, SE-97187 Luleå, Sweden. 3Faculty of Engineering, Al-Azhar University, P.O. Box 83513, Qena, Egypt. 4Faculty of Engineering, Minia University, Minia, Egypt, P.O. Box 61111, Minia, Egypt. Received: 19 September 2017 Accepted: 7 September 2018 References 1. T. Logeswari, M. Karnan, An improved implementation of brain tumor detection using segmentation based on hierarchical self organizing map. Int. J. Comput. Theory Eng. 2(4), 591 (2010) 2. E.-S. A. El-Dahshan, H. M. Mohsen, K. Revett, A.-B. M. Salem, Computer-aided diagnosis of human brain tumor through MRI: A survey and a new algorithm. Expert Syst. Appl. 41(11), 5526–5545 (2014) 3. D. Jayadevappa, S. Srinivas Kumar, D. Murty, Medical image segmentation algorithms using deformable models: a review. IETE Tech. Rev. 28(3), 248–255 (2011) 4."}], "source_title": "Two-phase multi-model automatic brain tumour diagnosis system from magnetic resonance images using convolutional neural networks", "source_doi": "10.1186/s13640-018-0332-4", "source_url": "https://jivp-eurasipjournals.springeropen.com/track/pdf/10.1186/s13640-018-0332-4", "source_license": "cc-by-sa", "zotero_item_key": "2ZZSU54R"}
{"id": "ae27d8583b760e5f", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Two-phase multi-model automatic brain tumour diagnosis system from magnetic resonance images using convolutional neural networks; DOI: 10.1186/s13640-018-0332-4; license: cc-by-sa."}, {"role": "assistant", "content": "S. Yazdani, R. Yusof, A. Karimian, M. Pashna, A. Hematian, Image segmentation methods and applications in MRI brain images. IETE Tech. Rev. 32(6), 413–427 (2015) 5. L. Zhao, K. Jia, Multiscale CNNs for brain tumor segmentation and diagnosis. Comput. Math. Methods Med. 2016, 7 (2016). https://doi.org/ 10.1155/2016/8356294. Article ID 8356294 6. A. Madabhushi, G. Lee, Image analysis and machine learning in digital pathology: challenges and opportunities. Med. Image Anal. 33, 170–175 (2016). 20th anniversary of the Medical Image Analysis journal (MedIA) 7. J. Jiang, P. Trundle, J. Ren, Medical image analysis with artificial neural networks. Comput. Med. Imaging Graph. 34(8), 617–631 (2010) 8. D. J. Hemanth, C. K. S. Vijila, A. I. Selvakumar, J. Anitha, Performance improved iteration-free artificial neural networks for abnormal magnetic resonance brain image classification. Neurocomputing. 130, 98–107 (2014) 9. G. E. Hinton, S. Osindero, Y.-W. Teh, A fast learning algorithm for deep belief nets. Neural Comput. 18(7), 1527–1554 (2006) 10. C. Yan, F. Coenen, B. Zhang, Driving posture recognition by convolutional neural networks. IET Comp. Vision. 10, 103–11411 (2016) 11. H.-I. Suk, S.-W. Lee, D. Shen, Deep ensemble learning of sparse regression models for brain disease diagnosis. Med. Image Anal. 37, 101–113 (2017) 12. A. Krizhevsky, I. Sutskever, G. E. Hinton, in Proceedings of the 25th International ConferenceonNeuralInformation Processing Systems – Volume 1.\n[page 10] Abd-Ellah et al. EURASIP Journal on Image and Video Processing (2018) 2018:97 Page 10 of 10 ImageNet Classification with Deep Convolutional Neural Networks (Curran Associates Inc., Lake Tahoe, Nevada, 2012), pp. 1097–1105 13. C. Yan, H. Xie, D. Yang, J. Yin, Y. Zhang, Q. Dai, Supervised hash coding with deep neural network for environment perception of intelligent vehicles. IEEE Trans. Intell. Transp. Syst. 19(1), 284–295 (2018). https://doi. org/10.1109/TITS.2017.2749965 14. A. ¸Senta¸s, ˙I. Tashiev, F. Küçükayvaz, S. Kul, S. Eken, A. Sayar, Y. Becerikli, in Advances in Internet, Data & Web Technologies, ed. by L. Barolli, F. Xhafa, N. Javaid, E. Spaho, and V. Kolici. Performance evaluation of support vector machine and convolutional neural network algorithms in real-time vehicle type classification (Springer, Cham, 2018), pp. 934–943 15. Y. LeCun, K. Kavukcuoglu, C. Farabet, in Proceedings of 2010 IEEE International Symposium on Circuits and Systems. Convolutional networks and applications in vision (IEEE, Paris, 2010), pp. 253–256 16. J. Lei, G. Li, J. Zhang, Q. Guo, D. Tu, Continuous action segmentation and recognition using hybrid convolutional neural network-hidden Markov model model. IET Comput. Vision. 10, 537–5447 (2016) 17. The Cancer Imaging Archive, RIDER NEURO MRI Database (2016). https:// wiki.cancerimagingarchive.net/display/Public/RIDER+NEURO+MRI. Accessed 14 Sept 2018 18. M. K. Abd-Ellah, A. I. Awad, A. A. M. Khalaf, H. F. A. Hamed, in Building Sustainable Health Ecosystems: 6th International Conference on Well-Being in the Information Society, WIS 2016, Tampere, Finland, September 16-18, 2016, Proceedings, ed. by H. Li, P. Nykänen, R. Suomi, N. Wickramasinghe, G. Widén, and M. Zhan. Classification of brain tumor MRIs using a kernel support vector machine, vol. 636 (Springer, Cham, 2016), pp. 151–160 19. M. K. Abd-Ellah, A. I. Awad, A. A. M. Khalaf, H. F. A. Hamed, in IEE International Conference on Microelectronics, ICM 2016, Giza, Egypt, December 17–20, 2016. Design and implementation of a computer-aided diagnosis system for brain tumor classification (IEEE, 2016), pp. 73–76 20. C. Lakshmi Devasena, M. Hemalatha, Efficient computer aided diagnosis of abnormal parts detection in magnetic resonance images using hybrid abnormality detection algorithm. Cent. Eur. J. Comput. Sci. 3(3), 117–128 (2013) 21. S. Patil, V. R. Udupi, A computer aided diagnostic system for classification of brain tumors using texture features and probabilistic neural network. Int. J. Comput. Sci. Eng. Inf. Technol. Res. 3, 61–66 (2013) 22. M. P. Arakeri, G. R. M. Reddy, Computer-aided diagnosis system for tissue characterization of brain tumor on magnetic resonance images. Sig. Image Video Process. 9(2), 409–425 (2015) 23. E. Dandıl, M. Çakıro˘glu, Z. Ek¸si, in ICT Innovations 2014. Computer-aided diagnosis of malign and benign brain tumors on MR images (Springer, Cham, 2015), pp. 157–166 24. S. Goswami, L. K. P."}], "source_title": "Two-phase multi-model automatic brain tumour diagnosis system from magnetic resonance images using convolutional neural networks", "source_doi": "10.1186/s13640-018-0332-4", "source_url": "https://jivp-eurasipjournals.springeropen.com/track/pdf/10.1186/s13640-018-0332-4", "source_license": "cc-by-sa", "zotero_item_key": "2ZZSU54R"}
{"id": "64372cbd4cafc5ca", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Intelligent Classification of Surrounding Rock of Tunnel Based on 10 Machine Learning Algorithms; DOI: 10.3390/app12052656; license: cc-by-sa."}, {"role": "assistant", "content": "Ensemble Learning Approach for the Prediction of Quantitative Rock Damage Using Various Acoustic Emission Parameters. Appl. Sci. 2021, 11, 4008. [CrossRef] 36. Han, M.-S.; Yang, I.-J.; Kim, K.-M. A prediction of the rock mass rating of tunnelling area using artificial neural networks. J. Korean Tunn. Undergr. Space Assoc. 2002, 4, 277–286. 37. Bressan, T.S.; de Souza, M.K.; Girelli, T.J.; Junior, F. Evaluation of machine learning methods for lithology classification using geophysical data. Comput. Geosci. 2020, 139, 104475. [CrossRef] 38. Sun, D.; Lonbani, M.; Askarian, B.; Armaghani, D.J.; Huynh, V.V. Investigating the Applications of Machine Learning Techniques to Predict the Rock Brittleness Index. Appl. Sci. 2020, 10, 1691. [CrossRef] 39. Schimazek, J.; Knatz, H. Der Einfluß des Gesteinsaufbaus auf die Schnittgeschwindigkeit und den Meißelverschleiß von Streckenvortriebsmaschinen. Glückauf 1970, 106, 274–278. 40. Schimazek, J.; Knatz, H. Die beurteilung der bearbeitbarkeit von gesteinen durch schneid-und rollenbohrwerkzeuge. Erzmetall 1976, 29, 113–119. 41. Gehring, K. Möglichkeiten zur Beurteilung des Arbeitsverhaltens von Werkzeuge zur schneidenden Gesteinsbearbeitung. Berg-Und Hüttenmännische Mon. 1973, 118, 319–327. 42. Becker, H.; Lemmes, F. Gesteinsphysikalische Untersuchungen im Streckenvortrieb. Tunnel 1984, 2, 71–76. 43. Thuro, K. Bohrbarkeit Beim Konventionellen Sprengvortrieb: Geologisch-Felsmechanische Untersuchungen Anhand Sieben Ausgewählter Tunnelprojekte; Inst. für Allg. u. Angewandte Geologie d. Ludwig-Maximilians-Univ.: Berlin, Germany, 1996. 44. Torno, S.; Torao, J.; Menéndez, M.; Gent, M.; Velasco, J. Mathematical and fuzzy logic models in prediction of geological and geomechanical properties of rock mass by excavation data on underground works. J. Civ. Eng. Manag. 2011, 17, 197–206. [CrossRef] 45. Mostofi, M.; Rasouli, V.; Mawuli, E. An estimation of rock strength using a drilling performance model: A case study in blacktip field, Australia. Rock Mech. Rock Eng. 2011, 44, 305. [CrossRef] 46. Naville, C.; Serbutoviez, S.; Throo, A.; Vincké, O.; Cecconi, F.J.O. Seismic while drilling (SWD) techniques with downhole measurements, introduced by IFP and its partners in 1990–2000. Oil Gas Sci. Technol. 2004, 59, 371–403. [CrossRef] 47. Kalantari, S.; Hashemolhosseini, H.; Baghbanan, A. Estimating rock strength parameters using drilling data. Int. J. Rock Mech. Min. 2018, 104, 45–52. [CrossRef] 48. Mehmet, C.; Osman, Y.A.; Serdar, Y. Relationships between the drilling rate index and physicomechanical rock properties. Bull. Eng. Geol. Environ. 2017, 76, 253–261. 49. Aydin, G.; Karakurt, I.; Hamzacebi, C. Performance Prediction of Diamond Sawblades Using Artificial Neural Network and Regression Analysis. Arab. J. Sci. Eng. 2015, 40, 2003–2012. [CrossRef] 50. Aydin, G.; Karakurt, I.; Hamzacebi, C. Artificial neural network and regression models for performance prediction of abrasive waterjet in rock cutting. Int. J. Adv. Manuf. Technol. 2014, 75, 1321–1330. [CrossRef] 51. He, M.; Li, N.; Yao, X.; Chen, Y. A New Method for Prediction of Rock Quality Designation in Borehole Using Energy of Rotary Drilling. Rock Mech. Rock Eng. 2020, 53, 3383–3394. [CrossRef] 52. Yue, Z.Q.; Lee, C.F.; Law, K.T.; Tham, L.G. Automatic monitoring of rotary-percussive drilling for ground characterization— Illustrated by a case example in Hong Kong. Int. J. Rock Mech. Min. 2004, 41, 573–612. [CrossRef] 53. Eldert, J.V.; Schunnesson, H.; Johansson, D.; Saiang, D. Application of measurement while drilling technology to predict rock mass quality and rock support for tunnelling. Rock Mech. Rock Eng. 2019, 53, 1349–1358. [CrossRef] 54. Galende-Hernández, M.; Menéndez, M.; Fuente, M.; Sainz-Palmero, G. Monitor-While-Drilling-based estimation of rock mass rating with computational intelligence: The case of tunnel excavation front. Automat. Constr. 2018, 93, 325–338. [CrossRef] 55. Chen, L.; Chen, S.; Tu, P.; He, T. Study on Mutual Relationships between Surrounding Rock Classifications by Q Value, RMR and BQ Method for Underground Cavern. Subgrade Eng. 2017, 6, 107–112. (In Chinese) 56. Aydin, G.; Karakurt, I.; Aydiner, K. Investigation of the surface roughness of rocks sawn by diamond sawblades. Int. J. Rock Mech. Min. 2013, 61, 171–182. [CrossRef]"}], "source_title": "Intelligent Classification of Surrounding Rock of Tunnel Based on 10 Machine Learning Algorithms", "source_doi": "10.3390/app12052656", "source_url": "https://www.mdpi.com/2076-3417/12/5/2656/pdf?version=1646383280", "source_license": "cc-by-sa", "zotero_item_key": "UGAERXYU"}
{"id": "0de67cb75047c774", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Spectral Angle Mapping and AI Methods Applied in Automatic Identification of Placer Deposit Magnetite Using Multispectral Camera Mounted on UAV; DOI: 10.3390/min12020268; license: cc-by-sa."}, {"role": "assistant", "content": "Best performing machine learning vs. deep learning classification models based on 0.12, 0.13 and 0.17 SAM cosine similarity thresholds at 2 m, 10 m and 20 m, respectively: (a) Comparisons based on global accuracy and pre-class precision; (b) Comparisons based on model training durations. Lastly, considering the UAV flight logs, as well as the AI model outputs in this study, one can come to the conclusion that the most effective UAV flight elevation is 20 m. At 20 m, one has a larger spatial area coverage, which eliminated the need to capture multitudes of images. Matching these variables with the attained AI model classifications, shows that the best classifier at this UAV flight elevation in identifying magnetite iron sands is the 1D CNN. 5. Conclusions This study proposed the employment of a 6-bands multispectral UAV drone-mounted camera in the automatic identification of magnetite iron sands at a placer deposit. As an initial step, the UAV drone was flown at three different flight elevations of 2 m, 10 m, and 20 m so as to find the optimum magnetite iron sand identification flight elevation. From the flight time, power consumption and number of spectral images capturing flight logs, 20 m flight elevation was the most viable. This study went on to examine the SAM analysis maps at these three flight elevations. It was found that the SAM cosine similarities could indeed identify the magnetite iron sand target areas at 2 m, 10 m and 20 m at 0.12, 0.13 and 0.17 cosine similarity threshold values, respectively, from the 6 bands multispectral images. Using the respective cosine similarity thresholds per flight elevation, the paper contin- ued to address the viability of coupling SAM analysis segmentation maps as well as ground truths, with different AI models. It was found that the proposed coupled system did indeed yield industry viable outputs across all flight heights. However, considering all outputs of this study, it was found that the best overall performance ratios were demonstrated at 20 m flight elevation. At this elevation, the UAV drone was able to capture 6-bands multispectral images with a spatial resolution of 338 m2. Furthermore, using the 6-bands multispectral imagery data, a 1D CNN DL model with a global accuracy of 99.7% and a per-class precision of 99.4% was found to be the overall best AI model. System capabilities with the potential for applicability in the mining industry/sphere have been demonstrated by this study. Said applications include rock and mineral ex- ploration from a remote location; mapping of mine sites via UAV drone technology; the ability to classify rocks and minerals based on their cosine similarities (SAM analysis) to a reference spectrum; the ability to take advantage of multispectral imaging from an elevation and distance of one’s choosing, and lastly, the flexibility of applying objective AI models in the classification of rocks and minerals (as demonstrated with magnetite).\n[page 18] Minerals 2022, 12, 268 18 of 19 Author Contributions: Conceptualisation, B.B.S. and Y.K.; Data curation, H.T.; Formal analysis, H.I.; Funding acquisition, Y.K. and T.A.; Investigation, B.B.S.; Methodology, B.B.S.; Resources, H.I.; Software, N.O.; Supervision, Y.K.; Validation, Z.B. and E.S.; Visualisation, H.I. and H.T.; Writing— original draft, B.B.S.; Writing—review & editing, Z.B., E.S., T.A. and Y.K. All authors have read and agreed to the published version of the manuscript. Funding: This work was supported by the Cooperative Research Project Program of the Penta-Ocean Construction Co., Ltd. Data Availability Statement: Not applicable. Acknowledgments: ‘Inter-University Exchange Project: An Innovative Program for Development of Core Human Resources for Smart Mining to Lead Sustainable Resource Development in Southern Africa’. Part of this work was carried out under the Cooperative Research Project Program of the Penta-Ocean Construction Co., Ltd. JSPS ‘Establishment of Research and Education Hub on Smart Mining for Sustainable Resource Development in Southern African Countries’ (JPJSCCB2018005). Conflicts of Interest: The authors declare no conflict of interest. Appendix A The procedure for coupling SAM with AI algorithms is as follows; At Figure 7 1⃝, a region within an image that is known to possess magnetite iron sands is pinpointed and used as a reference from which SAM is performed. As Figures 6 and 7 2⃝show, though a SAM may seem satisfactory from visual observations, it is rarely, if ever, 100% perfect."}], "source_title": "Spectral Angle Mapping and AI Methods Applied in Automatic Identification of Placer Deposit Magnetite Using Multispectral Camera Mounted on UAV", "source_doi": "10.3390/min12020268", "source_url": "https://www.mdpi.com/2075-163X/12/2/268/pdf?version=1645347690", "source_license": "cc-by-sa", "zotero_item_key": "VPIE7856"}
{"id": "80b1e2a26c9e5714", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Cutterhead and Cutting Tools Configurations in Coarse Grain Soils; DOI: 10.2174/1874836801711010182; license: cc-by."}, {"role": "assistant", "content": "The cutterhead selection and the cutting tools configuration are rather difficult in grasping the complex interaction between the pressurized shield driving and the ground. The complex interaction has been studied mainly by approaches of numerical simulations, laboratory tests and field observations. Manuel and Luis (2005) using discrete numerical models analyzed the thrust and torque needed for the shields to excavate the tunnels on two Madrid Metro extension projects, and studied soil stability problem at the tunnel face [1]. Shen et al. (2009) provided a 3D finite element model for dynamic simulation of the cutterhead and soil interaction in slurry shield tunneling by coupling the ALE approach in commercial package LS-DYNA with submodeling method in ANSYS [2]. Glenn and Mustafa (2011) analyzed critical operational characteristics during the design phase of shield machines with the use of Caterpillar proprietary discrete element method software, and examined how discrete element analysis could be applied to cutterhead performance optimization [3]. Hunt and Del Nero (2011) provided an overview of tunneling in the cobbly-bouldery ground, cited key references for more detailed study and gave emphasis to the key risk issues and approaches to mitigate them [4]. Lambrughi et al., (2012) developed a 3D numerical model using the software FLAC3D to simulate the overall process of excavation and construction of an EPB shield tunnel with the emphasis on the reasonable constitutive equation of soil [5]. Wu et al. (2013) presented a 3D model for cutterhead excavation system of an EPB machine with the discrete element method software PFC3D, and the measured system performance parameters indicating stability of excavation face, soil discharging rate, cutterhead system torque and cutter wear were in accordance with the in-situ data [6]. Katushin et al. (2013) briefly explained the importance and gave the insight on laboratory wear testing of the drag bit materials for tunneling machines [7]. Mosleh et al. (2013) presented an experimental study of tool wear using a laboratory tester which simulated the operating conditions of cutting tools in soil excavation operations, and analyzed effects of the relative hardness and moisture on tool wear in soil excavation operations [8]. Jakobsen et al. (2013) summarized the development of the Soft Ground Abrasion Tester (SGAT) apparatus, and showed its capabilities to evaluate, quantify and compare how the soil mineralogy, water content, pressure, compaction, and the use of soil conditioning additives influenced the wear rate on the SGAT excavation tool [9]. Nishitake (1987) described the EPB shields designed by Mitsubishi Heavy Industries, Ltd. to cope with large size boulders scattered in rather soft alluvial soil, which were equipped with both drag bits and rolling disc cutters on the cutterheads [10]. Dowden and Cass (1991) summarized the designs of shielded type machines and their application, and discussed the new developments which blend hard and soft ground tunneling technology [11]. Burger (2007) recalled evolution of the design of both slurry and EPB cutterheads, reviewed the state of the art, and highlighted both common as well as geological specific design features [12]. Herrenknecht and Bӓpler (2007) focused on the design and technical challenges of the three largest soft ground tunnel boring machines: two slurry shields (15.4 m diameter) for the Shanghai River Crossing project and one EPB shield (15.2 m diameter) for a road tunnel project in Madrid in Spain [13]. Burger and Wehrmeyer (2008a, 2008b) examined the development of the modern mixshield including the design of cutterhead, and looked at two particularly influential projects and their impact on the future mixshield technology [14, 15]. Metzger et al. (2009) focused on design considerations of the largest mixshield machine ever implemented in North America to excavate the 8534 linear m × 6.7 m ID tunnel for the City of Portland Oregon’s, Bureau of Environmental Services East Side CSO Tunnel Project, which included machine configuration, extended-wear protection, hyperbaric interventions, boulder digestion and so on [16]. Maidl and Comulada (2011) pointed out the key performance indicators for estimating performance of the EPB shields based on the evaluation of data from a great of projects [17]."}], "source_title": "Cutterhead and Cutting Tools Configurations in Coarse Grain Soils", "source_doi": "10.2174/1874836801711010182", "source_url": "https://openconstructionbuildingtechnologyjournal.com/VOLUME/11/PAGE/182/PDF/", "source_license": "cc-by", "zotero_item_key": "7KDIA2X2"}
{"id": "d6d6c67b104b885d", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Cutterhead and Cutting Tools Configurations in Coarse Grain Soils; DOI: 10.2174/1874836801711010182; license: cc-by."}, {"role": "assistant", "content": "On the other hand, reliable information about conditions of the cutting tools is an essential factor to minimize interventions to the necessary minimum and reduce risk of the extended compressed air work for repairing damaged tools and cutterheads. Using ultrasonic inspection equipment to detect wear of the cutting tools remains a promising approach with the ongoing development of the relevant technologies [22, 28]. CONCLUSION Based on the above statements and discussions, the following is concluded: The spoke type cutterheads equipped with precut bits, shell bits, scrapers and a center fish-tail bit are strongly 1. recommended for the shield tunneling in loose coarse grain soils without groundwater. The spoke-plate or plate type cutterhead with reasonable opening ratio is a better choice if the first priority is to 2. guarantee stability of the cutting face or mechanized driving will be performed in ground conditions rich in groundwater. The reasonable opening ratio of a cutterhead can be decided based on a good estimate of the size of particles in soils as well as experiences from the previous similar construction projects. Reasonable combinations of tools and proper height differences among tools for the specific particle size 3. distributions are of paramount importance to realize the efficient boring of tunnels in the coarse grain soils. Especially, the precut bits are suggested to be installed on both spoke type and plate type cutterhead on account of good performances revealed. The disc cutters, if mounted on cutterheads, will be single disc cutters with tungsten carbide inserts. Enough aggressive tools and robust cutterheads allowing for longer driving in the coarse grain soils are 4. suggested to be prepared or reserved to ensure the scheduled construction time lest the unforeseen scenarios or more difficult ground conditions will be encountered In the case that interventions and chamber access for changing tools are indispensable, the tools replacement 5. plan should be well planned beforehand and the operation procedure should be well prepared, as is helpful to reduce adverse effects caused by the over-sized particles on shield tunneling. CONFLICT OF INTEREST The authors confirm that this article content has no conflict of interest. ACKNOWLEDGEMENTS The authors gratefully acknowledge the financial support by the National Basic Research Program of China under Grant 2015CB057800. REFERENCES [1] M.J. Manuel, and L.E. Rodríguez, \"Discrete numerical model for analysis of earth pressure balance tunnel excavation\", J. Geotech. Geoenviron. Eng., vol. 131, pp. 1234-1242, 2005. [http://dx.doi.org/10.1061/(ASCE)1090-0241(2005)131:10(1234)] [2] J. Shen, X. Jin, Y. Li, and J. Wang, \"Numerical simulation of cutterhead and soil interaction in slurry shield tunneling\", Eng. Comput., vol. 26, pp. 985-1005, 2009. [http://dx.doi.org/10.1108/02644400910996862] [3] M. Glenn, and A. Mustafa, \"Discrete Element Method to predict soft ground cutterhead performance\", In: Rapid Excavation and Tunneling Conference Proceedings, 2011, pp. 1058-1067. [4] S.W. Hunt, and D.E. Del Nero, \"Tunneling in cobbles and boulders\", In: Breakthroughs in Tunneling Short Course, Colorado School of Mines, September 19-21, 2011. [5] A. Lambrughi, L. Medina, Rodríguez, and R. Castellanza, \"Development and validation of a 3D numerical model for TBM-EPB mechanised\n[page 18] 198 The Open Construction and Building Technology Journal, 2017, Volume 11 Li et al. excavations\", Comput. Geotech., vol. 40, pp. 97-113, 2012. [http://dx.doi.org/10.1016/j.compgeo.2011.10.004] [6] L. Wu, T. Guan, and L. Lei, \"Discrete element model for performance analysis of cutterhead excavation system of EPB machine\", Tunn. Undergr. Space Technol., vol. 37, pp. 37-44, 2013. [http://dx.doi.org/10.1016/j.tust.2013.03.003] [7] D. Katushin, M. Antonov, T.M. Vu, and D. Yung, \"Laboratory testing of materials for tunnel boring machine drag bits\", In: 3 th International Symposium PÄRNU, 2013, pp. 300-303. [8] M. Mosleh, E.A. Gharahbagh, and J. Rostami, \"Effects of relative hardness and moisture on tool wear in soil excavation operations\", Wear, vol. 302, pp. 1555-1559, 2013. [http://dx.doi.org/10.1016/j.wear.2012.11.041] [9] P.D. Jakobsen, L. Langmaack, F. Dahl, and T. Breivik, \"Development of the Soft Ground Abrasion Tester (SGAT) to predict TBM tool wear, torque and thrust\", Tunn. Undergr. Space Technol., vol. 38, pp. 398-408, 2013. [http://dx.doi.org/10.1016/j.tust.2013.07.021] [10] S."}], "source_title": "Cutterhead and Cutting Tools Configurations in Coarse Grain Soils", "source_doi": "10.2174/1874836801711010182", "source_url": "https://openconstructionbuildingtechnologyjournal.com/VOLUME/11/PAGE/182/PDF/", "source_license": "cc-by", "zotero_item_key": "7KDIA2X2"}
{"id": "2b7ca9bfd4703cb4", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Application of machine learning in predicting the rate-dependent compressive strength of rocks; DOI: 10.1016/j.jrmge.2022.01.008; license: cc-by-sa."}, {"role": "assistant", "content": "Therefore, the final output obtained from a new input vector x is: y ¼ 1 t X t l ¼ 1 ~yl ¼ 1 t X t l ¼ 1 ~hðx; ~SlÞ (13) An optimal RF-based model is realized by tuning the number of decision trees and the depth of each decision tree. The model can be more robust and accurate with a larger number of decision trees. However, the number is also associated with the calculation time, and the prediction accuracy is prone to stabilizing after hitting a critical point. The depth of the decision tree affects its own strength and the correlation between the trees. The purpose of the RF is to find a trade- off among the calculation time, decision tree strength, and correlation between the trees. When the terminal conditions are reached, the model is considered acceptable, and the training is completed. 3. Model training and validation We trained three machine learning models based on the SVM, the BPNN, and the RF algorithms by randomly selecting three quarters of the data in Table A1 and checked the performance of these models based on the remaining data. The compressive strength was the output, and the other parameters in Table A1 were used as the inputs (i.e. static compressive strength, P-wave velocity, specimen dimension, grain size, bulk density, and strain rate). We first qualitatively evaluated the performance of the three models based on the comparison between the predicted and the measured compressive strengths (Fig. 6). A data point on or close to the diagonal line indicates that the predicted compressive strength matches with the measured compressive strength. The data points with a small deviation from the diagonal line are mainly distributed in the region less than 120 MPa compressive strength. This is because the data points are densely distributed in this region, as shown in the histograms next to the upper and right axes. The results indicate that the amount of data points is critical for the accuracy of machine learning prediction. We then quantitatively compared the performance of these models based on the following metrics: MAE ¼ 1 n X n i ¼ 1 jyi \u0001 byij (14) MXAE ¼ maxjyi \u0001 byij (15) MAPE ¼ 1 n X n i ¼ 1 \u0005\u0005\u0005\u0005 yi \u0001 byi byi \u0005\u0005\u0005\u0005 \u0005 100% (16) Fig. 5. A schematic diagram of the principle of RF algorithm. M. Wei et al. / Journal of Rock Mechanics and Geotechnical Engineering 14 (2022) 1356e1365 1360\n[page 6] RMSE ¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi P n i ¼ 1 ðyi \u0001 byiÞ2 n v u u u t (17) where MAE, MXAE, MAPE, and RMSE denote the mean absolute er- ror, the maximum absolute error, the mean absolute percentage error, and the root mean square error, respectively; n represents the number of data points; and yi and byi are the predicted and the measured values for the ith strength value, respectively. The performance metrics plotted in Fig. 7 demonstrate that the compressive strengths predicted by the three machine learning models are mostly similar. The RF-based performance metrics are lower than the SVM- and BPNN-based performance metrics, manifesting the superiority of the RF over the SVM and the BPNN in predicting the compressive strength of rocks. The BPNN-based performance metrics are lower than the SVM-based performance metrics, except the MAPE. The high MAPE value is attributed to the significant absolute error obtained from the SVM-based model when predicting a high strength value. Given the better perfor- mance of the RF-based model, we used this model to investigate the rate dependence of rock strength in Section 4. 4. Discussion 4.1. Sensitivity analysis Our study demonstrates that the machine learning models perform well to predict the compressive strength of rocks using the six input parameters (i.e. static compressive strength, P-wave ve- locity, specimen dimension, grain size, bulk density, and strain rate). In rock engineering, one may not be able to collect sufficient data points to train the three models and to ensure the accuracy of machine learning prediction. Insufficient data points of single input parameter may influence significantly or negligibly the perfor- mance of these models. Here, we performed a sensitivity analysis to evaluate the importance of each input parameter in the strength prediction. In each model, we excluded one of the six input pa- rameters in turn. In other words, we inputted the other five pa- rameters in the model and compared the performance difference between the model with five input parameters and the model with six input parameters."}], "source_title": "Application of machine learning in predicting the rate-dependent compressive strength of rocks", "source_doi": "10.1016/j.jrmge.2022.01.008", "source_url": "https://www.sciencedirect.com/science/article/pii/S167477552200049X", "source_license": "cc-by-sa", "zotero_item_key": "KNCWQQJD"}
{"id": "6b1ee204a2ac64a6", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Application of machine learning in predicting the rate-dependent compressive strength of rocks; DOI: 10.1016/j.jrmge.2022.01.008; license: cc-by-sa."}, {"role": "assistant", "content": "Appl. Sci. 11 (19), 8973. Zhou, Y.X., Xia, K., Li, X.B., Li, H.B., Ma, G.W., Zhao, J., Zhou, Z.L., Dai, F., 2012. Sug- gested methods for determining the dynamic strength parameters and mode-I fracture toughness of rock materials. Int. J. Rock Mech. Min. Sci. 49, 105e112. Zhou, Z., Cai, X., Chen, L., Cao, W., Zhao, Y., Xiong, C., 2017. Influence of cyclic wetting and drying on physical and dynamic compressive properties of sandstone. Eng. Geol. 220, 1e12. Zhou, Z., Cai, X., Li, X., Cao, W., Du, X., 2020. Dynamic response and energy evolution of sandstone under coupled staticedynamic compression: insights from experimental study into deep rock engineering applications. Rock Mech. Rock Eng. 53, 1305e1331. Mingdong Wei is currently a research fellow in the School of Civil and Environmental Engineering, Nanyang Tech- nological University (NTU), Singapore. His research in- terests include rock fracture mechanics, rock dynamics, engineering geology, and cold regions geomechanics. He worked as a postdoctoral researcher at Aalto University, Finland, after obtaining his doctorate and bachelor’s de- gree from Sichuan University, China. He has authored or co-authored more than forty academic publications. He was awarded the Excellent Doctoral Dissertation Award 2018 by the Chinese Society for Rock Mechanics and Engi- neering, the Best Presentation Award 2021 (2nd runner- up) by the Society for Rock Mechanics & Engineering Geol- ogy, Singapore, the Research Scholarship Block (RSB)-Post- doctoral Fellowship 2019 by NTU, and the National Scholarship for Doctoral Students in China. M. Wei et al. / Journal of Rock Mechanics and Geotechnical Engineering 14 (2022) 1356e1365 1365"}], "source_title": "Application of machine learning in predicting the rate-dependent compressive strength of rocks", "source_doi": "10.1016/j.jrmge.2022.01.008", "source_url": "https://www.sciencedirect.com/science/article/pii/S167477552200049X", "source_license": "cc-by-sa", "zotero_item_key": "KNCWQQJD"}
{"id": "589ecd101a2099be", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Reinforcement learning based process optimization and strategy development in conventional tunneling; DOI: 10.1016/j.autcon.2021.103701; license: cc-by."}, {"role": "assistant", "content": "While [14], who trained their DQN to play classical Atari 2600 games, took a modified version of the original games’ scoring systems as a means to deal with the reward in each timestep, [15] only gave a reward of +1 or − 1 at the end of the game depending on whether or not the agent has won the respective Go-match. In our tunneling simulation, we designed a hierarchical point system that values (i.e. rewards/penalizes) either a state or an action. The rewarding/penalizing is done based on a list of conditions, that is worked through from top to bottom after every step of an episode and the agent is given the first reward where the condition is fulfilled. Except for the reward for achieving a breakthrough, all rewards are negative (i. Fig. 6. 3D visualization of the 2 × 2100 × 2 hypermatrix that represents the state of the simulation. G.H. Erharter et al.\n[page 7] Automation in Construction 127 (2021) 103701 7 e. penalties) and from top to bottom the penalties are sorted from biggest to smallest. It is therefore possible that multiple conditions are fulfilled in one state, but only the most severe penalty is given. Table 4 shows the list of rewards and their respective conditions. We designed the list of rewards based on practical engineering experience and requirements arising from training a completely unex­ perienced agent. For example, we penalize face instabilities higher than a too far distance between top heading and bench excavation, as the immediate safety hazard of an unstable face is bigger. In contrast to that, an even bigger penalty (i.e. -6 points) is triggered if the agent would try to do bench excavation ahead of the top heading excavation. A geo­ technician in real world tunneling would know from the start that doing so brings numerous technical-, logistical- and safety problems, but an untrained RL-agent does not. During the experiments for this study we tried to design the reward system in a way that it is the sum of all penalties that would be fulfilled in a given state and not a hierarchical system. However, doing this was not beneficial for training and seemed to confuse the agent as it apparently did not know what it was punished for. For example, a combination of the penalty for a too far distance between top heading and bench (i.e. -3) plus the penalty for using face support (i.e. -2) would sum up to −5 which is the same penalty as the one for unstable face conditions. As given in Section 2.1, the evaluation to check if a newly excavated face is stable or not, is done based on the face pressure equation for open face tunneling from [29] (Eq. (3)). In Eq. (3), pf is the required pressure to achieve stable face conditions, where a pf < 0 indicates stable con­ ditions and a pf ≥0 indicates unstable conditions, respectively the amount of pressure that is necessary to stabilize the face. γR is the ground’s unit weight, D the (equivalent) tunnel diameter (see chapter 3.1), d the advance length (describing the unsupported area of the un­ lined wall), c′ the effective cohesion and φ′ the effective friction angle. pf = γRD* ⎛ ⎜ ⎜ ⎜ ⎝ 2 + 3* ( d D )6*tanφ ′ 18*tanφ ′ −0.05 ⎞ ⎟ ⎟ ⎟ ⎠− c ′ tanφ ′ (3) 4. Training In RL, one episode is the whole succession of states in between an initial and a terminal state [26]. In other words, an episode is one whole match of a game, or in this simulation, one whole sequence of actions that ultimately should lead to a breakthrough of the tunnel. We started training with an ε of 1 (i.e. “initial exploration” in Table 2) to promote exploration in the initial phase of training. The exploration decay of 0.99997 (see Table 2) that decreases ε following Eq. (2) was determined by trial and error. A smaller exploration decay (i.e. faster reaching of the final exploration) has shown to increase in­ stabilities in the training process at an early stage which are presenting themselves in spontaneous increases of the loss and decreases of the reward. With the given exploration decay, the final exploration is reached after 99,858 episodes of training. After this point ε is kept at a constant value of 0.05. Training does not need to be aborted after reaching the final exploration. In the current simulation, we aimed at training the agents for 120,000 episodes to observe one full epsilon decay and some episodes beyond that to check for stable conditions in different rockmasses, as illustrated in Fig. 7 (except for cases where training became unstable at some point; see next section)."}], "source_title": "Reinforcement learning based process optimization and strategy development in conventional tunneling", "source_doi": "10.1016/j.autcon.2021.103701", "source_url": "https://www.sciencedirect.com/science/article/pii/S0926580521001527", "source_license": "cc-by", "zotero_item_key": "RXHUVV3X"}
{"id": "cd48a5cf38af5be5", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Two-phase multi-model automatic brain tumour diagnosis system from magnetic resonance images using convolutional neural networks; DOI: 10.1186/s13640-018-0332-4; license: cc-by-sa."}, {"role": "assistant", "content": "4.2 Accuracy evaluation criteria To evaluate the two system phases, the following stan- dard metrics were used: sensitivity, which represents the proportion of actual positives that are correctly classified; specificity, which indicates the proportion of negatives Table 7 Performance analysis of the developed localization R-CNN-based method compared to other methods in the literature Localization method Performance evaluation metrics (%) Layers Sensitivity PPV DICE Havaei et al. [29] 10 87.00 88.00 0.88 Chang [30] 22 – – 0.87 Casamitjana et al. [31] 22 83.08 84.60 0.84 Zhao et al. [32] 16 83.00 92.00 0.87 Pereira et al. [33] 51 93.00 80.00 0.85 Pereira et al. [34] 11 89.00 88.00 0.83 Proposed method 05 95.44 81.80 0.87\n[page 8] Abd-Ellah et al. EURASIP Journal on Image and Video Processing (2018) 2018:97 Page 8 of 10 that are correctly classified; and accuracy, which is the proportion of both true positives and true negatives. These three metrics were calculated using Eqs. 8, 9, and 10, respectively [18, 19]: Sensitivity = σ σ + \u0004 (8) Specificity = η η + \u0006 (9) Accuracy = σ + η σ + η + \u0004 + \u0006 (10) where true positives (σ) are the correctly classified pos- itive cases, true negatives (η) are the correctly classified negative cases, false positives (\u0006) are the incorrectly classified positive cases, and false negatives (\u0004) are the incorrectly classified negative cases. In addition to sensitivity, accuracy, and specificity, other parameters have been considered for evaluating tumour detection and localization phases. The new parame- ters were balanced accuracy (BA) = (sensitivity + speci- ficity)/2, positive predictive value (PPV) = σ/(σ + \u0006), and negative predictive value (NPV) = η/(η + \u0004) [11]. Addi- tionally, DICE scores are considered a common parameter for tumour localization approaches. 4.3 Tumour detection accuracy analysis The performance of the CNN is measured in every net- work layer. The goal of this step is to identify the optimum number of network layers for achieving the highest accu- racy and the shortest feature extraction time. The mea- sured performance of each layer is recorded in Table 3. It is demonstrated in Table 3 that setting the number of layers equal to 9 provided the best performance. Table 4 shows the best accuracy obtained from the three CNN models used in this study and the corresponding layer number. The network size has been optimized to give the best performance in learning phase and then has been fixed, with the validation and testing phases using the same network size as the learning phase. This is to generalize the CNN model for all data samples. To further investigate the performance of classical learning models against the proposed deep CNN-based model, a classical brain tumour diagnosis method has been applied and studied using the same database con- figurations as in Table 2. The proposed CNN system achieved 99.55% accuracy, while the classical method accomplished 66.96%. A short comparison between the performance of the previous classical method and the newly proposed deep learning system using the same database is presented in Table 5. The obtained results in Table 5 show the superiority of the deep learning method in terms of brain tumour detection accuracy, which is reflected in the MRI classification. The proposed CNN-based system has achieved 99.55% accuracy for tumour diagnosis when selecting nine CNN layers. To ensure the credibility of the obtained results, Fig. 4 Visual results in the axial view from the localization phase. The top row shows the T2 modality. The second row shows the ground truth. The bottom row shows the segmented images\n[page 9] Abd-Ellah et al. EURASIP Journal on Image and Video Processing (2018) 2018:97 Page 9 of 10 a comparison with results from the literature was per- formed. The outcome of the proposed CNN-based system compared with those of three popular methods reported in the literature is reported in Table 6. 4.4 Tumour localization accuracy analysis Simulation experiments have been conducted to show the performance of the proposed system in fulfiling the tumour localization task. A comparison with other studies found in the literature is shown in Table 7. It is clear that the proposed system is superior to the others in terms of the network size, sensitivity and DICE metrics but not the PPV metric. A sample of visual results from the MRIs database used in the localization phase of the experimental work is shown in Fig. 4."}], "source_title": "Two-phase multi-model automatic brain tumour diagnosis system from magnetic resonance images using convolutional neural networks", "source_doi": "10.1186/s13640-018-0332-4", "source_url": "https://jivp-eurasipjournals.springeropen.com/track/pdf/10.1186/s13640-018-0332-4", "source_license": "cc-by-sa", "zotero_item_key": "2ZZSU54R"}
{"id": "f04982580f385f26", "messages": [{"role": "system", "content": "You are Gemma4Tunnel, a careful tunnel boring machine and tunnelling engineering assistant. Use engineering terminology precisely, state assumptions, and avoid inventing source claims."}, {"role": "user", "content": "Extract reusable tunnelling/TBM engineering knowledge from this allowed source excerpt. Preserve important parameters, mechanisms, risks, and caveats. Do not add unsupported claims.\n\nSource: Cutterhead and Cutting Tools Configurations in Coarse Grain Soils; DOI: 10.2174/1874836801711010182; license: cc-by."}, {"role": "assistant", "content": "Among others, Le (2003) recommended the key technical parameters concerning configurations of the shield cutterhead and cutting tools based on analyzing typical ground conditions encountered in Beijing area [24]; Wang (2005) introduced the first application of the EPB tunneling method in Beijing subway construction to dig Line 5 tunnels and described the main design parameters of the shield machine [25]; Liu and Song (2007) dealt with the problem of cutting tools wear in building Chengdu metro Line 1 tunnels in coarse grain soils and presented some recommendations after investigating failure causes of the disc cutters and scrapers [26]; Huang (2008) discussed influences of the oversized grains in the water-rich sandy gravel soils of Lot 13 of Shenyang metro Line 2 on configurations of the cutterhead and cutting tools as well as layout of the screw conveyor [27]; Zhang et cl. (2008) presented the monitored abrasions of cutting tools used in building some shield tunnel of Beijing subway line 4, and established the relation between abrasion coefficients of tools and shield tunneling parameters [28]; Wang (2008) analyzed the crucial technical parameters, such as opening ratio of the cutterhead and combination of the cutting tools fit for Chengdu geological conditions relying on the experiences and lessons from the construction of Phase 1 projects of Chengdu metro Line 1 [29]; Wang and Zhang (2011) studied the arrangement and structure of the cutterhead and cutting tools suitable for boring subway tunnels in Shenyang medium-coarse sand stratum in virtue of the experiences from construction of the tunnels on Shenyang metro Line 1 [30]; Zhu (2011) briefly illuminated the use of the spoke type cutterhead with opening ratio of 60% and the combination of precut bits and face scrapers in the Phase 2 Lot 7 project of Beijing subway Line 10 [31]. Nearly all known publications in China concentrated on only a single case study mainly because of the tribal knowledge of the involved tunnel project. Generalized studies from the cases study in combination with some undocumented recordings are necessary to improve the shield tunneling work, resulting in a more economic process in the coarse grain soils and further providing references for the succeeding similar projects. Due to the theoretical difficulties in mastering the interactions between coarse grain soils and cutting tools, the common understanding and recognition reached from the in-situ performances of the shield cutterheads remain the well accepted approach for configurations of the cutterhead and cutting tools suitable to dig tunnels in the coarse grain soils. On the basis of briefly summarizing main characteristics of the coarse grain soils disclosed in China, the conclusions on which consensus can be reached are generalized from the typical construction cases, and the rule-of-thumb suggestions on the configurations of the cutterhead and cutting tools are put forward. 2. MAIN CHARACTERISTICS OF COARSE GRAIN SOILS ENCOUNTERED IN CHINA It is of paramount importance to grasp mechanical and physical properties of the coarse grain soils to be encountered when the EPB shield tunneling method will be utilized. Presented in Fig. (1) is the typical sandy gravel/cobble soil with boulders found in building Beijing subway tunnels. Main characteristics of the coarse grain soils disclosed in China subway tunnel construction can be briefly summarized as follows: Typical physical and mechanical properties are the large content of 30-70% by volume of coarse grains (gravel, 1. cobble, and boulder), the general grain size of 30-200 mm with the maximum size of over 500 mm and the uniaxial compressive strength of grains as high as 180 MPa, taking the coarse soils encountered in Beijing for instance. Particularly, in the tunnel project of Lot 3 of Beijing subway Line 9, the content of the coarse grains is as high as 80%, and the maximum particle size is more than 1000 mm. The coarse grain soils are of unfavorable plasticity and fluidity, which will bring serious abrasions to shield 2. machines."}], "source_title": "Cutterhead and Cutting Tools Configurations in Coarse Grain Soils", "source_doi": "10.2174/1874836801711010182", "source_url": "https://openconstructionbuildingtechnologyjournal.com/VOLUME/11/PAGE/182/PDF/", "source_license": "cc-by", "zotero_item_key": "7KDIA2X2"}