Remaining useful life

Estimates how many cycles a turbofan engine has left, from its sensor history. Built by Alvin Alias.

RMSE 15.85 / Ridge 17.47 direct local comparison on the FD001 test set

AA-2026-01 server waking, first run can take a minute

Picks one of 10 held-out engine sequences from NASA's simulated FD001 test set. The official answer appears only after the model commits.

Same columns as the sample file. download the sample CSV

Run log
Sensor traces

Calibration record · direct comparison separated from literature context
RMSE on the FD001 official test set
15.85this model, local evaluation
17.47Ridge baseline, valid direct local comparison
16.14Zheng 2017 reported value, context only
18.45Babu 2016 reported value, context only

100 engines run to failure · 20,631 training cycles · 14 informative sensors after dropping 7 flat channels

The literature protocols have not been audited for parity with this repository. Only the local Ridge result is a direct comparison; the published values are context, not evidence that this model beat them.

How it works · features, split, model

I split by engine, never at random. A random split drops cycles from the same engine into both train and test, so the model learns that engine's own drift instead of the shape of degradation.

The features are 30-cycle rolling means and spreads. Thirty cycles is a chosen smoothing horizon, not a measured fouling period or a validated optimum.

The largest SHAP attribution is associated with the rolling mean of HPC outlet temperature. That is consistent with a degradation interpretation, but SHAP describes this model's associations and does not establish physical cause.

Notes · what this can and cannot tell you
  1. This ML demo sleeps after extended inactivity. First wake can take a moment; runs after that are quick.
  2. The model was trained with remaining life capped at 125 cycles, the standard convention for this dataset, so no estimate can exceed 125.
  3. The range shown is a ±15 cycle heuristic, not a formal prediction interval.
  4. The known records are held-out engine sequences from NASA's simulated FD001 dataset. Their official answers ship with this page and stay hidden until after each run.
  5. One engine's estimate says nothing about a different engine type or operating condition; FD001 is a single operating condition.