AI4I 2020 machine-failure model (HistGradientBoosting, ONNX)
Predicts whether a milling machine fails on the next cycle from six process readings, using the configuration of smart-factory-app: raw features plus three domain features (mechanical load = torque × rpm, thermal stress, tool-wear / torque), HistGradientBoosting with the repository's tuned parameters.
Try it: huggingface.co/spaces/senanurcetin/ai4i-failure-risk (runs in the browser).
Holdout performance (80/20 stratified, seed 42, 68 failures in the holdout)
| metric | value |
|---|---|
| ROC-AUC | 0.987 |
| PR-AUC (the metric that matters at 3.4% failures) | 0.901 |
| precision / recall at 0.5 | 0.930 / 0.779 |
Across five independent splits the repository reports PR-AUC 0.880 ± 0.032: with ~68 failures per holdout, a single split's PR-AUC moves by several points. Random failures (RNF) are essentially unpredictable, and tool-wear failures are only partly captured.
Inputs
Float32 vector in this order (also in model.json): type (L=0, M=1, H=2), air_temp_k,
process_temp_k, rpm, torque_nm, tool_wear_min, mechanical_load, thermal_stress,
tool_wear_load_ratio. Output: label and class probabilities (index 1 = failure).
ONNX vs scikit-learn: max probability difference 1e-7 on the holdout. train.py reproduces it.
Data
AI4I 2020 Predictive Maintenance Dataset (UCI, doi:10.24432/C5HS5C), CC BY 4.0. A synthetic dataset modelled on real milling data; numbers here do not transfer to a specific machine without retraining.