Failure check · exploratory cost-scenario threshold

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

Air temp (K) 290 to 310 K Process temp (K) 305 to 315 K Speed (rpm) 1168 to 2886 rpm Torque (Nm) 3.8 to 76.6 Nm Tool wear (min) 0 to 253 min

waiting

failure probability

waiting for prediction

Why the threshold is 0.775

Using a modeled $50,000 miss and $2,000 false alarm, a 25:1 ratio, the held-out count sweep reaches its minimum at 0.775. The same test labels were used to select and report this threshold, so threshold-dependent metrics are exploratory. If misses and false alarms count equally, the grid minimum is about 0.95 (note 2).

What moved this prediction

Calibration record · what makes 0.775 exploratory

Exploratory recall 91.2% at 0.775	62 of 68 test-set failures caught, 6 missed, 83 false alarms; the same labels selected the threshold
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, so the model uses SMOTE plus class weighting. A separate sensitivity sweep applies assumed miss and alarm costs to the held-out counts. Because that same test set selects and reports the 0.775 threshold, the threshold-dependent results are exploratory. SHAP describes model associations, not physical causes.

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 assumptions, not measured customer costs. Their 25:1 ratio gives a test-count minimum at 0.775; equal costs give about 0.95.
The same held-out labels selected the threshold and produced its recall and confusion counts. Treat every threshold-dependent metric as exploratory until a separate validation split or nested evaluation is used.
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.