Work

Cross-Bearing Remaining Useful Life Under Dataset Shift: An Audited Comparison on PRONOSTIA and XJTU-SY

Year: 2026
Type: article
Abstract: Cross-bearing remaining useful life (RUL) assessment can be misleading when serial vibration records are treated as independent units. We retrospectively compared Ridge-based predictors on six PRONOSTIA development bearings under nested leave-one-bearing-out assessment, then evaluated source-frozen Baseline Ridge and State-Gated Ridge on 15 XJTU-SY bearing runs (8,766 predictions). The simple reference attained the lowest development macro RMSE (1.4319 h). In the exploratory cross-rig analysis, baseline and gated macro RMSE were 4.820294 and 4.835116 h, respectively; gating improved seven bearings. A paired 5,000-draw bearing-bootstrap 95% interval for gated-minus-baseline RMSE was [−0.043242, 0.075461] h. Target-bearing recordings supplied local signal normalisation without target RUL labels or model retraining. The last-recording endpoint was a proxy rather than independently verified failure. Archived predictions and scoring were independently audited, but physical sensor-response equivalence was not established.
Language: English
Open Access status: diamond
OpenAlex ID: W7237364914
FWCI:
Cites: 15
Cited by:
Related to: