Silent Sensor Faults in Remaining Useful Life Prediction: A Robustness Benchmark, Uncertainty Audit and Model-Agnostic Input Repair
Remaining useful life (RUL) models are evaluated mainly on clean sensor data, whereas deployed monitoring systems lose channels and receive plausible but erroneous values when a sensor freezes, drifts, becomes noisy or spikes. Nine model families were evaluated on the four Commercial Modular Aero-Propulsion System Simulation (C-MAPSS) subsets, and five of them on PHM08, under a deterministic protocol of 24 corruption scenarios, one fault family withheld from training and the clean condition. The root mean square error (RMSE) increased on average by up to 24% under single-sensor drift and by 49–121% under unseen spikes. The robustness of a gated recurrent encoder was traced to the training regime. The coverage of conformalised 90% prediction intervals fell to 70–82% under spikes while the mean interval width of the deep models changed by less than 3%. Since missing entries were tolerated far better than spiked or drifting ones, a detect-then-mask repair was developed that masks entries flagged by clean-data statistics. Applied to 360 saved C-MAPSS checkpoints without retraining, it reduced RMSE under unseen spikes by 30% and under strong drift by 26% at a clean-data cost of 0.04 cycles, and with Mondrian recalibration it restored the coverage under spikes to 89%.
Authors
- Cansu Akyürek Anacur (ORCID: https://orcid.org/0009-0005-8608-7892)
- Özlem Şeker (ORCID: https://orcid.org/0000-0002-8686-340X)
Institutions
- Bitlis Eren University (TR)
Publication Details
- Journal
- Sensors
- Published
- 2026-11-18
- DOI
- https://doi.org/10.3390/s26227290
- Primary Topic
- Machine Fault Diagnosis Techniques
- Type
- article
- Field-Weighted Citation Impact
- 0.00