| --- |
| license: mit |
| task_categories: |
| - question-answering |
| pretty_name: EngineMT-QA |
| size_categories: |
| - 100K<n<1M |
| --- |
| |
|
|
| # EngineMT-QA Dataset |
|
|
| ## Overview |
|
|
| **EngineMT-QA** is a large-scale, multi-task, multimodal dataset for **Time-Series Question Answering (Time-Series QA)**. It enables research on aligning multivariate time-series signals with natural language through four key cognitive tasks: |
|
|
| * **Understanding** |
| * **Perception** |
| * **Reasoning** |
| * **Decision-Making** |
|
|
| The dataset is built on **N-CMAPSS**, simulating real-world aero-engine operational and maintenance scenarios. It supports the development and evaluation of temporal-textual models that integrate sensor signals and natural language queries. |
|
|
| ## Dataset Contents |
|
|
| * `time_series_data.h5` (18GB): Multivariate time-series data with 600 time steps × 33 sensor channels |
| * `train_qa.jsonl` (39MB): 87k training QA pairs |
| * `test_qa.jsonl` (16MB): 25k test QA pairs |
|
|
| Total: \~113k QA instances across four tasks. |
|
|
| ## Tasks and Annotations |
|
|
| Each QA sample is aligned to specific time series segments (IDs match `data_ID` in `.h5`), and annotated for: |
|
|
| | Task | Input Scope | Format | Evaluation Metrics | |
| | --------------- | ------------ | --------------- | ------------------ | |
| | Understanding | Single-cycle | Open-ended | BLEU, ROUGE-L | |
| | Perception | Single-cycle | Multiple-choice | Accuracy, F1 | |
| | Reasoning | Multi-cycle | Multiple-choice | Accuracy, F1 | |
| | Decision-Making | Multi-cycle | Open-ended | BLEU, ROUGE-L | |
|
|
| Each `conversations` entry contains multi-turn Q\&A annotated with: |
|
|
| * `stage`: Task type (1=Understanding, 2=Perception, 3=Reasoning, 4=Decision) |
| * `attribute`: Question format (`open` or `close`) |
| * `value`: Natural language prompt or response |
|
|
| ## Example QA Entry |
|
|
| ```json |
| { |
| "id": ["81571"], |
| "name": ["81571.csv"], |
| "conversations": [ |
| { |
| "from": "human", |
| "stage": "3", |
| "attribute": "close", |
| "value": "Given the time series signal <ts>, by perceiving the engine signal across 10 cycles ... what is the qualitative condition of the engine?\n a: Good Condition\n b: Moderate Condition\n ..." |
| }, |
| { |
| "from": "gpt", |
| "value": "d" |
| } |
| ] |
| } |
| ``` |
|
|
| ## How to Use |
|
|
| ### Load Time-Series Data |
|
|
| ```python |
| import h5py |
| with h5py.File('time_series_data.h5', 'r') as f: |
| data_ids = f['data_ID'][:] |
| sequences = f['seq_data'][:] # Shape: (118921, 600, 33) |
| ``` |
|
|
| ### Load QA Data |
|
|
| ```python |
| import json |
| with open('train_qa.jsonl', 'r') as f: |
| for line in f: |
| sample = json.loads(line) |
| ``` |
|
|
| ### Match QA with Time-Series |
|
|
| The `id` field corresponds to indices in the `.h5` file. Use this to align QA queries with sensor sequences for modeling or fine-tuning. |
|
|
| ## Applications |
|
|
| * Time-Series QA benchmarking |
| * Engine health diagnostics |
| * Fault detection and prognosis |
| * Multimodal model training and evaluation |
| * Cross-modal semantic alignment (text + sensor signals) |
|
|
| ## Construction Details |
|
|
| * Based on NASA's **N-CMAPSS** dataset. |
| * Questions designed using domain-specific rules and LLMs (e.g., ChatGPT, DeepSeek) refined by experts. |
| * QA tasks were segmented to reflect real-world cognition: |
|
|
| * **Understanding**: Signal meaning interpretation |
| * **Perception**: Health/fault recognition |
| * **Reasoning**: Degradation trend & RUL |
| * **Decision-Making**: Maintenance strategy suggestions |
|
|
| More details can be found in **Appendix A–C** of the associated paper. |
|
|
| ## Citation |
|
|
| If you use this dataset, please cite our ICML 2025 paper: |
|
|
| ```bibtex |
| @inproceedings{wang2025itformer, |
| title = {ITFormer: Bridging Time Series and Natural Language for Multi-Modal QA with Large-Scale Multitask Dataset}, |
| author = {Yilin Wang and Peixuan Lei and Jie Song and Haoyu Zhe and Tao Chen and Yuxuan Zhang and Lei Jia and Yuanxiang Li and Zhongyu Wei}, |
| booktitle = {International Conference on Machine Learning (ICML)}, |
| year = {2025} |
| url ={https://huggingface.co/papers/2506.20093} |
| } |
| ``` |
|
|
| ## License |
|
|
| \[Insert license type here – e.g., MIT, CC BY-NC-SA 4.0, etc.] |
|
|
| ## Contact |
|
|
| For questions or collaborations, please contact: |
| [pandalin@sjtu.edu.cn] |
