BearLLM: A Prior Knowledge-Enhanced Bearing Health Management Framework with Unified Vibration Signal Representation
English | įŽäŊä¸æ
## đĨ NEWS - **[2025-04-11]** đ The [AAAI-25 Proceedings](https://aaai.org/proceeding/aaai-39-2025/) are now officially published! Our [conference paper](https://ojs.aaai.org/index.php/AAAI/article/view/34188) is included. We welcome you to read and cite it! - **[2025-03-06]** đ The complete dataset and code are now officially open source! - **[2024-12-11]** âĢ We are now working on making the code of BearLLM public. Stay tuned! - **[2024-12-10]** đ The BearLLM paper is accepted by the Thirty-Ninth AAAI Conference on Artificial Intelligence ([AAAI-25](https://aaai.org/conference/aaai/aaai-25/)). - **[2024-08-21]** đ The preprint of the BearLLM paper is available on arXiv. Check the [paper page](https://arxiv.org/abs/2408.11281) for more details. ## đ TODO - [ ] Improve related comments and documentation. - [x] Upload the complete BearLLM demo code. - [x] Upload the health management corpus of the MBHM dataset. - [x] Collect the codes for pre-training and fine-tuning BearLLM. - [x] Collect the codes of BearLLM's classification network and other comparison models. - [x] Upload the vibration signal portion of the MBHM dataset. ## đ Introduction The [MBHM](https://huggingface.co/datasets/SIA-IDE/MBHM) dataset is the first multimodal dataset designed for the study of bearing health management. It is divided into two parts: vibration signals and health management corpus. The vibration signals and condition information are derived from 9 publicly available datasets, and are still under continuous updating and improvement. The thousands of working conditions pose more difficult challenges for the identification model and better represent real-world usage scenarios. [BearLLM](https://github.com/SIA-IDE/BearLLM) is a prior knowledge-enhanced bearing health management framework with a unified vibration signal representation. This framework transforms the signal to be tested into the frequency domain, enabling effective identification of spectral differences compared to the vibration signal under fault-free conditions. By aligning the vibration signal with the fault semantic embedding, we achieve a unified natural language response for various health management tasks through a fine-tuned language model with low computational overhead. Experiments demonstrate that this framework achieves leading performance under thousands of working conditions. ## đģ Requirements The code is implemented in Python 3.12. The required packages are listed in the `requirements.txt` file. You can install the required packages by running the following command: ```bash conda create --name bearllm python=3.12 conda activate bearllm pip install -r requirements.txt ``` ## đ Quick Start ### 1. Download Demo Data / Use Your Own Data First, you need to download the `demo_data.json` from the [MBHM](https://huggingface.co/datasets/SIA-IDE/MBHM/tree/main) dataset. For users in mainland China, you can use the [mirror link](https://hf-mirror.com/datasets/SIA-IDE/MBHM/tree/main) to speed up the download: Or, you can also build your own test data in the same format: `instruction`: Text instruction for health management task. `vib_data`: Vibration signal data to be identified, with a required duration of 1 second. `ref_data`: Reference vibration signal data without faults, with a required duration of 1 second. ```json { "instruction": "xxx.", "vib_data": [1.0, 0.0, 1.0, ...], "ref_data": [1.0, 0.0, 1.0, ...], } ``` ### 2. Download Weights You can download the pre-trained weights of [Qwen2.5-1.5B](https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct/tree/main) from Hugging Face. Additionally, you need to download the weights of [BearLLM](https://huggingface.co/SIA-IDE/BearLLM/tree/main). ### 3. Organize Files It is recommended to organize the weights and test data as follows: ``` BearLLM/ âââ qwen_weights/ â âââ model.safetensors â âââ tokenizer.json â âââ config.json â âââ other files... âââ bearllm_weights/ â âââ vibration_adapter.pth â âââ adapter_config.json â âââ adapter_model.safetensors âââ mbhm_dataset/ âââ demo_data.json ``` ### 4. Run Code First, copy the `.env.example` file to `.env` and modify the data paths inside. Then, you can run the code using the following command: ```bash python run_demo.py ``` ## âī¸ Development ### 1. Download Dataset First, you need to download the following files from the [MBHM](https://huggingface.co/datasets/SIA-IDE/MBHM/tree/main) dataset. For users in mainland China, you can use the [mirror link](https://hf-mirror.com/datasets/SIA-IDE/MBHM/tree/main) to speed up the download: - `data.hdf5`: Contains the vibration signal data. - `corpus.json`: Contains the health management corpus. - `metadata.sqlite`: Contains metadata information of the dataset. ### 2. Download Weights You can download the pre-trained weights of [Qwen2.5-1.5B](https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct/tree/main) from Hugging Face. ### 3. Modify Environment Variables Copy the `.env.example` file to `.env` and modify the data paths inside. ### 4. Pre-train and Fine-tune Model Pre-train according to `src/pre_training.py`. Fine-tune according to `src/fine_tuning.py`. ## đ Citation Please cite the following paper if you use this study in your research: ``` @article{pengBearLLMPriorKnowledgeEnhanced2025, title = {{{BearLLM}}: {{A Prior Knowledge-Enhanced Bearing Health Management Framework}} with {{Unified Vibration Signal Representation}}}, author = {Peng, Haotian and Liu, Jiawei and Du, Jinsong and Gao, Jie and Wang, Wei}, year = {2025}, month = apr, journal = {Proceedings of the AAAI Conference on Artificial Intelligence}, volume = {39}, number = {19}, pages = {19866--19874}, issn = {2374-3468}, doi = {10.1609/aaai.v39i19.34188}, urldate = {2025-04-11}, } ```