Turbofan Remaining Useful Life — Case Study

HEALTHY

CYCLES LEFT

Engine demo — RUL 95

failure (0) now: 95 cycles left healthy (125+)

Red line = the maintenance threshold (20 cycles). Act while the bar is still above it.

drag to explore

See it: the engine's remaining life, live

This is the heart of the RUL Explorer, running right here on the page — no video, no image, just inline SVG. Drag the slider (or press Play degradation) and watch the engine age: the fan slows, the combustor reddens, and the health verdict flips HEALTHY → CAUTION → CRITICAL as the remaining cycles drain toward the maintenance line. "This engine is running out of life" becomes something you see.

Role

Solo — data engineering, ML, visual application

Dataset

NASA C-MAPSS turbofan (FD001–FD004, run-to-failure)

Stack

Python, pandas, scikit-learn, XGBoost, LightGBM, CatBoost, Streamlit

Best so far

FD001 Extra Trees — RMSE 16.96 · FD002 Random Forest — RMSE 16.14

Status

Research — all four fleets benchmarked (FD001–FD004)

A jet engine does not fail without warning — it degrades, cycle by cycle, and its sensors carry the signature of that decline. The question this project answers is the one a maintenance planner actually asks: how many cycles of useful life are left before this engine crosses into unsafe territory? Not a stop/go flag, but a countdown — so an engine can be pulled for service while it is still healthy, rather than discovered too late.

The data

Source: NASA's C-MAPSS (Commercial Modular Aero-Propulsion System Simulation) turbofan degradation benchmark — the standard dataset for prognostics research. It ships as four sub-datasets that get progressively harder, differing in how many operating conditions the engines run under and how many fault modes can develop:

Fleet	Operating conditions	Fault modes	Why it's harder
FD001	1	1	Cleanest — one regime, one failure path
FD002	6	1	Six regimes muddy the sensor signal
FD003	1	2	Two competing degradation modes
FD004	6	2	Hardest — six regimes and two faults

Each engine is a run-to-failure trajectory of 21 sensor channels and 3 operating settings; the target is the remaining useful life at every point along the way. Following standard practice, RUL is piecewise-linear capped (an engine far from failure is simply "healthy", so the label is clipped at a ceiling of ~125 cycles) — degradation is only meaningfully predictable once wear actually begins.

Eight regressors, four separate leaderboards

Every fleet was benchmarked with the same eight regressors, each tuned by a genuine hyperparameter search inside the reusable AEGIS model-execution framework: Ridge, HistGradientBoosting, LightGBM, XGBoost, CatBoost, Random Forest, Extra Trees, and SVR. Two deliberate discipline choices shape how the results are read:

Never pooled — one leaderboard per fleet. FD001 and FD002 are different physical problems (one regime versus six), so mixing their errors into a single ranking would be meaningless. Each sub-dataset gets its own RMSE-ranked leaderboard, and the champion is named per fleet — not a single flattering average.

Ranked on RMSE and the NASA prognostics score. RMSE measures raw accuracy, but the NASA score is asymmetric — it punishes a late prediction (saying an engine has more life than it does) far more harshly than an early one, because in the real world a late call is the dangerous one. Both are reported, so a model that is accurate and fails safe is visible.

FD001 — the clean regime (7 of 8 models)

One operating condition, one fault mode. The tree ensembles cluster tightly at the top; Extra Trees leads. (SVR is O(n²) and still completing on the full search — reported when done rather than rushed.)

Rank	Model	RMSE	MAE	R²	NASA score
1	Extra Trees	16.96	11.85	0.821	823.9
2	LightGBM	17.16	11.92	0.817	802.0
3	CatBoost	17.16	12.00	0.817	889.0
4	Random Forest	17.21	12.06	0.816	859.5
5	XGBoost	17.32	11.95	0.813	887.2
6	HistGradientBoosting	17.62	12.30	0.807	864.5
7	Ridge (linear baseline)	20.88	16.47	0.728	1312.2

FD002 — six operating regimes (7 of 7 models)

Six operating conditions make the sensor signal far noisier — yet the gradient-boosted and tree ensembles actually score a lower RMSE here than on FD001, because there is far more data to learn the regimes from. Random Forest takes it, with CatBoost a whisker behind. (SVR is skipped on the six-regime sets by design — it is prohibitively slow on ~50–60k rows; all seven applicable models completed.)

Rank	Model	RMSE	MAE	R²	NASA score
1	Random Forest	16.14	11.99	0.859	1586.7
2	CatBoost	16.19	11.88	0.858	1586.6
3	Extra Trees	16.21	12.13	0.857	1679.0
4	XGBoost	16.24	11.88	0.857	1612.6
5	LightGBM	16.33	11.89	0.855	1619.0
6	HistGradientBoosting	16.47	11.95	0.853	1635.8
7	Ridge (linear baseline)	18.15	14.07	0.821	1589.9

FD003 — two fault modes, one operating condition

One operating condition but two competing degradation modes (HPC + Fan). The tree ensembles lead again and cluster tightly at the top — Random Forest takes it, Extra Trees a hair behind. All eight models completed.

Rank	Model	RMSE	MAE	R²	NASA score
1	Random Forest	18.93	13.33	0.767	1601
2	Extra Trees	19.06	13.26	0.763	1784
3	CatBoost	19.57	14.04	0.750	1750
4	HistGradientBoosting	19.85	13.80	0.743	1834
5	LightGBM	19.92	13.97	0.741	1830
6	XGBoost	20.05	14.26	0.738	1970
7	Ridge (linear baseline)	20.73	15.88	0.720	1671
8	Support Vector Regressor	22.30	15.11	0.676	3160

FD004 — the hardest fleet (5 of 7 models)

Six operating regimes and two fault modes — the toughest board of all. In the deep FULL search, Random Forest and Extra Trees did not finish (Random Forest was pulled for being pathologically slow on this set), so among the completed models CatBoost leads. The complete short-run benchmark crowns Random Forest (RMSE 18.68) as FD004's true champion — noted honestly rather than hidden.

Rank	Model	RMSE	MAE	R²	NASA score
1	CatBoost	18.88	13.79	0.807	2733
2	HistGradientBoosting	19.32	13.99	0.798	2967
3	XGBoost	19.40	14.13	0.796	3025
4	LightGBM	19.43	14.12	0.796	3164
5	Ridge (linear baseline)	20.85	16.56	0.765	2404

Why the visual matters

A prognostics model's output is a number, and numbers are easy to nod at and ignore. Binding that number to a living engine — colour, spin, glow, a draining bar against a hard maintenance line — turns "RUL = 18" into an unmistakable "pull this engine now." The point of the whole project is pre-emptive action, and the interface is built to make that action feel urgent at exactly the right moment, and calm when there is still runway.

Engineering rigor

Separate leaderboards, honestly named. Four fleets, four rankings — no pooling of physically different problems into one flattering average.

The right cost function. The NASA asymmetric score is reported alongside RMSE precisely because, for a jet engine, a prediction that is late is far more dangerous than one that is early.

The linear baseline stays in. Ridge regression is kept in last place on every board rather than dropped — an honest benchmark shows how far a simple model gets, and how much the ensembles actually earn.

Resume-safe by design. The framework checkpoints every model, so a multi-day run survives interruption — exactly how FD004's deep search was managed. All four fleets are now benchmarked; FD004 reports the 5 of 7 models that finished, with the unfinished two flagged rather than estimated.