Bearing fault diagnosis based on multiscale subdomain full-dimensional dynamic convolutional networks

Abstract

To address the issues of poor data distribution alignment and suboptimal transfer performance in existing transfer learning methods for unsupervised bearing fault diagnosis, this paper proposes a Multi-Scale Sub-Domain Full-Dimensional Dynamic Convolution Network (MSODNet) for bearing fault diagnosis. First, Omni-dimensional Dynamic Convolution (ODConv) is introduced to replace the conventional CNN convolution layers, leveraging its dynamic kernel weight adjustment capability to capture more comprehensive fault features. Second, the Local Maximum Mean Discrepancy (LMMD) is employed to refine the Maximum Mean Discrepancy (MMD) by incorporating class information, enabling fine-grained alignment of sub-distributions for the same category across source and target domains. Finally, the ECANet module is integrated into the model to perform channel-wise adaptive calibration of extracted features, enhancing the model’s focus on critical features, improving feature representation, and reducing noise interference. Experimental results demonstrate that the proposed method effectively captures both the divergence and commonality of signals under multi-condition scenarios, achieving an average accuracy of 99.35% in bearing vibration fault diagnosis. This confirms the superior diagnostic performance and robustness of the MSODNet model.