Bearing Fault Diagnosis via Graph Autoencoder Networks with Multi-kernel Subdomain Adversarial Domain Adaptation

Unsupervised domain adaptation (UDA) based on transfer learning methods have been widely used in the research of bearing fault diagnosis under variable operating conditions, and some useful results have been achieved. However, conventional UDA methods predominantly prioritize the extraction of class labels and domain labels from the data, neglecting the effect of data architecture information on extracted characteristics. In addition, global domain adaptation methods ignore the relationship between subdomains. Therefore, in this paper, we propose multi-kernel subdomain adversarial domain adaptation for graph autoencoder networks (MSADAGAE) to solve the above problems, which has two key parts. Firstly, multiple graph convolutional blocks are used to obtain graph node features as well as topologies at different scales via residual-connected graph autoencoders (GAEs). Second, a subadaptation module based on multi-layer multi-kernel local maximum mean discrepancy (MLMK-LMMD) is proposed, including a globally aligned domain classifier and subdomain-aligned domain adaptation. Then, the optimization of feature classification boundaries is further enhanced through margin loss regularization. Finally, validation is performed on the public datasets CWRU, JNU, and the results show that the model exhibits good performance even under unbalanced datasets.