Abstract

Substantial progress has been made in bearing-fault classification under same-condition settings, yet cross-condition diagnosis remains affected by three coupled issues: speed-induced temporal-scale perturbation, insufficient use of structured label information, and domain shift across operating conditions. To address these issues, this paper presents BearingPRO, a unified framework that combines physics-guided representation alignment with group ordinal labeling for bearing-fault classification. The framework contains three modules. First, a TimeWarp module applies controlled temporal stretching and compression to emulate waveform variations induced by speed changes. Second, a Grouped Ordinal module introduces intra-group ordinal constraints according to the hierarchical relation between fault type and fault severity. Third, a Physics-Guided Representation Alignment (PGRA) module uses rotational-speed priors for carrier-frequency calibration, envelope extraction, and cross-domain alignment. On the CWRU bearing dataset, under the 0 HP → 3 HP transfer task, BearingPRO achieves 0.9205 ± 0.0088 accuracy and 0.8962 ± 0.0105 Macro-F1 in the unified reproduction setting used in this study. Relative to the re-implemented comparison methods under the same backbone and training budget, the proposed framework yields higher mean performance and lower variance. Ablation results further indicate that temporal-scale modeling, grouped ordinal supervision, and physics-guided alignment play complementary roles in the current setting. At the same time, the scope of the conclusion is explicitly bounded: the present evidence is obtained on the CWRU test rig, within the considered speed range, and under artificially introduced point defects; therefore, the method is presented as a well-supported cross-condition classifier for this benchmark, not as a universally validated solution for all motors.

IPC Classification

Keywords