A rolling bearing fault diagnosis framework based on multi-modal feature fusion and marginal-conditional alignment

Abstract As a key component in rotating machinery, rolling bearings frequently operate under variable speeds and loads. The diagnostic performance of fault classification models is severely constrained due to significant differences in the distribution of operational data under variable working conditions. Although transfer learning effectively mitigates performance degradation caused by such variations through aligning cross-domain feature distributions, methods relying solely on single-modal features struggle to comprehensively capture the multi-dimensional features of fault information, leading to a decline in the generalization ability of the model. To address this issue, a rolling bearing fault diagnosis framework named multi-modal feature fusion marginal-conditional alignment (MMFF-MCA)is proposed in this study, integrating a MMFF method and a MCA strategy. Specifically, the MMFF extracts key features from time–frequency images and time-domain signals through a linear deformable star network and a temporal network (TimesNet), respectively, and employs a bidirectional dynamic gated attention fusion mechanism for adaptively weighted fusion of multi-scale features. The MCA combines multi-kernel maximum mean discrepancy (MK-MMD) with contrastive conditional alignment to enhance cross-domain feature consistency and refine fault state discrimination. Cross-domain transfer experiments on two bearing datasets demonstrate that the MMFF-MCA framework outperforms single-modal and existing fusion methods in diagnostic performance, validating its effectiveness for fault diagnosis in complex industrial scenarios.