Lead time — the business-facing chart
Recall tells an engineer nothing they can act on. Lead time is how long they have between the alert and the machine stopping, and it is the only number that decides whether a reactive repair becomes a planned intervention.
threshold —
Is there anything to detect? The run-up, measured
Causal aggregates over the seven days before each failure started. This is the evidence a predictive model has available, and it is not the same for every event.
| Event | Start | Duty, −168 h | Duty, −72 h | Oil °C, −168 h | Motor A, −168 h |
|---|---|---|---|---|---|
| F1 | 2020-04-18 00:00:00 | 0.160 | 0.068 | 57.4 | 1.71 |
| F2 | 2020-05-29 23:30:00 | 0.113 | 0.087 | 66.2 | 1.72 |
| F3 | 2020-06-05 10:00:00 | 0.169 | 0.139 | 65.0 | 2.07 |
| F4 | 2020-07-15 14:30:00 | 0.209 | 0.250 | 68.7 | 2.81 |
Read the rows against each other
F4 has a textbook run-up. F2 and F3 have almost no aggregate precursor, so a model that
"catches" them is reacting to the event, not anticipating it. F1 runs backwards — duty
falls to zero because the train was parked — and a careless model learns "idle means
failure imminent", which is spurious and will fire every night in 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.