Predictive maintenance · C-MAPSS FD001 turbofan fleet · 100 held-out engines · 21 sensors (14 informative) · single operating condition · 1 cycle = 1 flight
RUL FORECAST ERROR
10.13 cyc
GBDT test MAE · RMSE 13.54 · 100 held-out engines
NEAR-FAILURE CAUGHT
21 / 25
84% recall at ≤ 30-cycle dispatch threshold
ALERT PRECISION
95%
22 alerts issued · 1 false alarm · 4 missed
INFORMATIVE SENSORS
14 / 21
7 constant channels screened out
FLEET HEALTH — DEGRADATION CURVESdescriptive index · s11·s4·s2 · train engines
— ENG-001 · 192 flights— ENG-060 · 172 flights— ENG-100 · 200 flights— ENG-034 · 195 flights
CYCLES TO FAILURE — 100 TRAIN ENGINESmedian 199 flights
MAINTENANCE QUEUE — PREDICTED RUL, TEST ENGINES (top 12)GBDT · dispatch threshold ≤ 30 cycles ≈ flights · bar = predicted, ghost = true
PREDICTION ERROR BY LIFE-STAGE (GBDT · test engines)RMSE cycles
Most accurate exactly where it matters: engines within 30 flights of failure (MAE 4.08 cycles). Mid-life is hardest — the piecewise 125-cycle cap leaves genuine ambiguity.
MODEL COMPARISON — TEST RMSE (cycles)100 held-out engines
GBDT on engineered window statistics wins; the raw-window transformer beats ridge but not the engineered features at 100 training engines — the same classical-vs-attention finding as the wind, solar and grid notebooks.
TOP CONTRIBUTING SIGNALS
Degradation *rate* (slope over the last 30 flights) of the turbine sensors dominates — the model is reading wear speed, not just level.
DISPATCH SIMULATION — ALERT AT ≤ 30 CYCLES
| model | alerts | TP | FP | missed | prec | recall |
|---|---|---|---|---|---|---|
| GBDT | 22 | 21 | 1 | 4 | 96% | 84% |
| transformer | 23 | 22 | 1 | 3 | 96% | 88% |
| ridge | 15 | 15 | 0 | 10 | 100% | 60% |
Of 25 test engines that truly fail within 30 flights, GBDT catches 21 with 1 false alarm; the slightly higher recall of the transformer (88%) is a calibration artifact, not a modeling win.
Real predictive-maintenance data — NASA C-MAPSS FD001 turbofan degradation (100 run-to-failure engines · 21 sensors) · GBDT RUL model trained on 20,631 flight cycles · AssetIQ runs on the telemetry you already collect