Interpretable Bearing Fault Diagnosis Using Synchronized Acoustic and Rotational-Speed Sensing Under Variable-Speed Conditions

Abstract

Variable-speed: operation causes fixed-duration acoustic frames to cover different shaft angles, reducing the mechanical consistency of bearing-fault features. This study proposes OSAP-KANFormer, an interpretable diagnostic framework using synchronized acoustic and rotational-speed measurements. Equal-angle resampling establishes a common mechanical scale, while a Hilbert-initialized learnable analytic layer and reciprocal cross-attention model amplitude–phase information. The final decision passes through predicted physical concepts and a sparse additive Kolmogorov–Arnold classifier. Under source-record-separated evaluation, the method achieved a Macro-F1 of 97.3% on a fixed-speed benchmark and 92.4% in leave-one-speed-out evaluation for previously observed bearing entities. Compared with ASFE-Transformer, Macro-F1 increased by 7.3 percentage points under unseen speeds and by 9.2 percentage points in constant-speed-to-variable-speed transfer. A separately trained external-dataset evaluation achieved 87.0% under a compound speed–load–microphone shift. Eleven physical concepts obtained a mean prediction R2 of 0.891. The results support the contribution of mechanical-angle alignment to variable-speed diagnosis and provide an explicit concept-level decision pathway. Applicability remains dependent on speed-reference quality, concept coverage, and the evaluated transfer setting.