A Standardized and Multi-Perspective Benchmark for Predictive Maintenance on Multi-Sensor Industrial Time-Series Data
We recommend Python 3.8+ and a virtual environment. Then run:
pip install -r requirements.txtDownload and unzip the datasets into the ./dataset/ directory. We support all standard datasets defined in the benchmark, including:
- Paderborn, CWRU, FEMTO, IMS, MFPT, XJTU, HUST Bearing, etc.
Our dataset can be downloaded from https://huggingface.co/submission096
See the dataset_mapping inside the codebase for full descriptions.
To train and evaluate a model via CLI:
python run.py --task_name classification --model TimesNet --data 01 --is_training 1Modify --model, --data, and other hyperparameters as needed. All models listed in TSLib (e.g., TimeMixer, iTransformer, PatchTST, Koopa, etc.) are supported.
To launch the interactive web application:
streamlit run app.pyThis interface allows you to:
- Browse and visualize time-domain and frequency-domain characteristics of PdM datasets
- Select, configure, and train SOTA models interactively
- View and compare model performance across datasets via dynamic leaderboards
- Monitor training progress and visualize metrics in real time
Figure 1: Dataset selection and time-/frequency-domain feature visualization.
Figure 2: Model configuration, training progress, and real-time performance feedback.
Figure 3: Cross-dataset performance analysis and leaderboard comparison of models.
As shown in Figure 1, users can interactively explore raw time-series signals along with their frequency-domain representations to understand sensor characteristics and class distributions.
Figure 2 demonstrates the model training module, where users can configure model parameters and monitor training and validation metrics in real time.
Finally, Figure 3 presents the benchmarking interface, where metrics such as accuracy, F1 score, calibration error, and inference time are summarized across datasets. This helps practitioners select and tune models based on task-specific requirements.