C-MAPSS RUL benchmark

The remaining-useful-life (RUL) prognostics benchmark over NASA’s public C-MAPSS turbofan degradation dataset (FD001–FD004). The harness, metric suite (RMSE, PHM08 asymmetric score, α-accuracy), CLI (python -m vela.eval rul), and a synthetic fixture are built and unit-tested; this page captures the real-data run once the dataset is present.

Method

An interpretable single-sensor degradation floor: pick the most monotonically-degrading sensor, project its trend to the failure threshold, and score the resulting RUL estimates with the standard PHM metrics. It is deliberately simple — the honest, interpretable baseline that a KG-prior or frozen-TSFM RUL head (RAP-440 / RAP-441) must beat, not a state-of-the-art claim.

Reproduce

make data-cmapss          # fetch FD001–004 (tries mirrors) → data/cmapss/ + MANIFEST.sha256
                          # or drop the {train,test,RUL}_FD00x.txt files in manually
make bench-rul            # python -m vela.eval rul --dataset all → results/bench/rul_cmapss.json

The dataset bytes are pinned by data/cmapss/MANIFEST.sha256, so the metrics below are exactly reproducible from a byte-identical dataset (the RUL harness is pure-stdlib and deterministic — no RNG, no wall-clock in the payload).

Results

Pending data acquisition

The real FD001–004 run has not been captured in this environment (the dataset is an external NASA download; data/cmapss/ is empty here). Run the two commands above on a network-connected machine and paste the observed PHM table here — as observed, never hardcoded (guardrail #2). Until then this benchmark is roadmap, not shipped, and no numbers are quoted.

Dataset	RMSE	PHM08	α-accuracy	n units
FD001	—	—	—	—
FD002	—	—	—	—
FD003	—	—	—	—
FD004	—	—	—	—