Maintenance decisions for degrading bearings depend on whether an asset can complete the next planned operating interval. The study assesses mission-window survival probabilities and their suitability for maintenance decisions.
XJTU-SY provides the primary run-to-failure data, and PRONOSTIA provides external experimental validation. Threshold, linear, ensemble, history-aware and calibrated models are compared under bearing- and condition-disjoint splits using discrimination, calibration and decision-risk measures.
Advanced and calibrated variants improve selected cases but do not uniformly outperform random forest or root-mean-square thresholds; model rankings also vary by evaluation criterion.
The validation uses experimental run-to-failure bearing datasets rather than industrial field deployment. PRONOSTIA provides external experimental validation, but horizon units are acquisition file indices and may not have the same physical-time meaning as XJTU-SY.
Calibrated probabilities can support continued operation, inspection or intervention after site-specific validation of horizons, thresholds and failure consequences.
The study unifies horizon-specific reliability targets, disjoint shift evaluation, probability calibration and maintenance-facing risk metrics rather than proposing a new predictor.
