Choosing an operating point with the room

Accuracy is discarded. Two numbers replace it: a false-alarm budget the service organisation can absorb, and the recall and lead time achievable inside that budget. Move the budget and watch what the model can actually deliver.

Model

False alarms per operating day: 1.00

Events caught

–

of 4 documented

Mean lead time

–

hours before the failure started

Window recall

–

the conventional metric

Alert episodes

–

true / false

Watch the two metrics disagree Window recall stays close to nothing while events caught reaches three of four. They are measuring different things. An engineer is not paid to classify ten-minute windows; they are paid to arrive before the machine stops. Episode-level detection with a stated lead time is the metric that survives contact with the depot.

Computed by scripts/build_models.py in 86.4s at 2026-08-09T19:08:15+00:00 · scikit-learn 1.6.1 · seed 20260809 · source: MetroPT-3 (Air Production Unit of a Metro do Porto train), CC BY 4.0, DOI 10.24432/C5VW3R. Nothing on this page is hardcoded.