Read the signals.
Validate the horizon.

A tuned gradient-boosting regressor on rolling-window sensor features substantially beats a naive median-RUL baseline on NASA's own held-out test protocol, using a grouped cross-validation strategy that AI4I's static dataset structure cannot support.

No real plant sensors or live engine deployment are connected to this analysis.

IMPROVEMENT ON BASELINE63.8%Reduction in test RMSE

PHM08 SCORE837.6Asymmetric late-error penalty

TRAIN / TEST ENGINES100 / 100NASA FD001 (one operating condition, one fault mode)

01 / DEGRADATION INTELLIGENCE

From sensor history to life estimates.

Analyze the degradation trajectory before inspecting model diagnostics and tuning details.

Why this exists alongside the AI4I case study

AI4I 2020 is a static snapshot dataset — one row per independently sampled machine, no time axis — so genuine remaining-useful-life regression isn't possible there. C-MAPSS is a real run-to-failure dataset, which is.
Because engines are tracked over many cycles, rows can be grouped by engine for cross-validation — the grouped/time-aware split AI4I's structure could not support is used here.
Evaluation uses the dataset's own designed protocol (NASA's official test split + truthed RUL labels), not a custom holdout.

Model vs. Naive Baseline

Approach	RMSE	PHM08 Score
Naive: always predict median training RUL (103.0)	49.8199	166570.5426
HistGradientBoostingRegressor (rolling-window features, tuned)	18.0535	837.6035

Model Performance

Sample engine degradation trajectories across cycles — Sample engine degradation trajectories

RMSE comparison: tuned model vs naive median baseline — Model vs. baseline RMSE

Predicted vs actual RUL scatter on the official test set — Predicted vs. actual RUL

Hyperparameter Tuning

Search method	RandomizedSearchCV (15 iter)
CV strategy	GroupKFold grouped by unit_number — rows from the same engine never appear in both the train and validation side of a fold. This is the grouped/time-aware split AI4I's static, one-row-per-machine structure cannot support.
Scoring	neg_root_mean_squared_error
Best CV RMSE	18.2949

Rolling-Window Feature Impact

Feature set	RMSE	PHM08 Score
Raw sensors only	18.1154	944.7926
+ Rolling window (mean/std)	18.0535	837.6035

Signal Importance (SHAP)

sensor_4_roll_mean 9.7938

sensor_9_roll_mean 6.3002

sensor_11_roll_mean 5.6202

sensor_15_roll_mean 4.7687

sensor_3_roll_mean 2.3657

sensor_14_roll_mean 2.224

sensor_2_roll_mean 2.197

sensor_21_roll_mean 1.3719

sensor_20_roll_mean 1.0545

Sensors

Total sensors	21
Kept (informative)	15
Dropped (near-constant)	sensor_1, sensor_10, sensor_16, sensor_18, sensor_19, sensor_5

Mean absolute SHAP value ranking of RUL model features — Mean |SHAP| feature importance

Limitations and next step

FD001 is the simplest C-MAPSS subset (one operating condition, one fault mode) — FD002/FD004 (six conditions, two fault modes) are harder and not covered here.
The RUL cap (125 cycles) follows common practice but is a modeling choice, not a physical constant.
Per-cycle 'true RUL' shown in sample trajectory charts before an engine's final test cycle is a linear back-projection from the single officially truthed point, not an independently verified value.
This is a simulated benchmark (PHM08 challenge data), not flight data from a real fleet.