A Learnable Amplitude-Phase Spectral Learning Framework for Cross-Domain Bearing Fault Diagnosis

Monitoring the health of critical rotating components is essential to ensure equipment safety and reduce economic losses. Although data-driven fault diagnosis methods have achieved promising results, their performance often deteriorates when applied across different domains due to distributional shifts. To address this challenge, a novel learnable amplitude-phase spectral learning framework is proposed for cross-domain bearing fault diagnosis. Unlike conventional methods that rely solely on amplitude spectra while ignoring phase information, the proposed framework introduces a learnable amplitude-phase adjuster that jointly refines both magnitude and phase components in the Fourier domain. Neglecting phase information may lead to the loss of essential time-domain structure and dynamic characteristics, thereby impairing the model’s ability to generalize across domains. This mechanism enhances the model’s ability to extract discriminative and domain-invariant features from vibration signals. Furthermore, a global-local maximum mean discrepancy (GL-MMD)-based alignment strategy is incorporated to reduce feature distribution discrepancies at both global and class-conditional levels. Extensive experiments conducted on two benchmark bearing datasets under varying working conditions demonstrate the effectiveness and strong generalization ability of the proposed method in cross-domain scenarios. On the triaxial bearing dataset, AP-GLANet achieves 95.29% average accuracy, and on CWRU it reaches 99.98% with 11/12 transfers attaining 100%.