Machine-health findings

Findings in plain language

Summary

We built a health-monitoring baseline from 100 engines (20,631 rows of cycle-by-cycle sensor data) without using any failure labels. On a held-out set of 30 engines the model separated healthy-period readings from end-of-life readings almost perfectly (AUC 0.99), and its health score rose toward end of life in every one of those 30 engines. This is a strong result, with the important caveat that it rests on the assumption that each engine record runs all the way to failure.

What we did

We trained an unsupervised model on a healthy window only — the earliest portion of each training engine's life (30% healthy fraction) — so the model learned what "normal" looks like without ever being told which engines fail or when. No labels were used in training. We then measured a "surprise" score (how far current readings drift from healthy behavior, taken as the maximum across the thermal, pressure, rotation, flow, and bleed sensor domains) and checked whether that score behaves sensibly on engines the model never saw during training. The 100 engines were split into 70 for training and 30 for holdout evaluation.

What the data shows

• Healthy vs. end-of-life separation (holdout): AUC of 0.99. In plain terms, if you pick a healthy-period reading and an end-of-life reading at random, the model ranks the end-of-life one as more degraded almost every time.
• Trend with age (holdout): In all 30 holdout engines the health score trended upward over the engine's life (fraction of units with positive correlation = 1.0). The average rank correlation with age was 0.708 and the median was 0.76. Individual engines ranged from weaker trends (e.g. 0.317, 0.366, 0.369, 0.409) to strong ones (e.g. 0.881, 0.889, 0.896).
• Statistical check: A permutation test gave a mean p-value of 0.0, meaning these trends are very unlikely to be chance.
• When the signal appears: The surprise score stays low and flat through most of life and climbs sharply near the end. Across life-tenths the holdout values move from about 4.13 early up through 9.803, 24.936, 95.478, and 389.421 in the final tenth. This tells you the warning shows up late but clearly — useful for catching degradation, but not necessarily far in advance.
• Consistency: The in-sample training numbers (AUC 0.992, mean correlation 0.728) are close to the holdout numbers, which is reassuring — the model is not just memorizing the training engines.

Caveats

• Holdout size: Results are based on 30 held-out engines. That is a reasonable check but a modest sample.
• Run-to-failure assumption: The protocol assumes each engine record runs all the way to failure. Because this file has no explicit failure or RUL label (confirmed in the warnings — check whether labels are stored separately), the trend-vs-age results should be treated as exploratory. The "surprise rises near end of life" finding only holds if the last cycles truly are end-of-life.
• Late warning: The surprise score rises steeply only in the final life-tenths. This baseline is better at confirming late-stage degradation than at giving long lead time.
• Weak-signal channels: Several channels (T2, P2, epr, farB, and the excluded tra_set, Nf_dmd, PCNfR_dmd) are near-constant and contribute little. This is expected for this type of data and does not undermine the result, but it means the useful signal comes from a subset of sensors.
• Domain placement: htBleed was placed in its own Bleed domain; a reviewer may prefer to merge it into thermal or pressure. This is a modeling choice worth a quick confirmation.

Suggested next step

Confirm where the failure/RUL labels for these engines are stored, then re-check the model's health score against a handful of known failures. This is low effort and would convert the current exploratory trend result into a validated one before any operational use.