Operating point
Draws a real record from the AI4I test data. What actually happened to the machine stays hidden until the model commits.
Set the values yourself
waiting
failure probability
waiting for prediction
Why the threshold is 0.775
The default 0.5 threshold treats a $50,000 miss and a $2,000 false alarm as equal mistakes. Sweeping the threshold against those costs puts the optimum at 0.775 (note 2).
What moved this prediction
What moved this prediction
Calibration record · why 0.775 and not 0.5, in dollars
| Recall 91.2% at threshold 0.775 | 62 of 68 test-set failures caught, 6 missed, 83 false alarms |
|---|---|
| PR-AUC 0.841 | XGBoost with SMOTE and class weighting |
| ROC-AUC 0.979 | held-out test set |
| Baselines | random forest PR-AUC 0.820, logistic regression 0.455 |
| Class balance | Failures are 3.4% of records; accuracy is meaningless here and is not reported |
How it works · imbalance, SMOTE, threshold sweep
Failures are rare and expensive, so the work is in the imbalance and the economics: SMOTE plus class weighting for recall, then a threshold sweep that minimizes modeled total cost instead of maximizing F1; the strongest SHAP factors are torque, wear, and speed, which matches the mechanics.
Notes · the costs are modeled and stated in the open
- This ML demo sleeps after extended inactivity. First wake can take a moment; runs after that are quick.
- The $50,000 and $2,000 figures are modeled costs, chosen to represent unplanned downtime against a scheduled maintenance call; change the costs and the optimal threshold moves. The sweep is in the repo.
- Known records are drawn from the AI4I 2020 dataset with their recorded outcomes; misses and false alarms print as plainly as catches.
- This model classifies near-term failure; the turbofan page estimates remaining life. Same domain, different question.