Domain adaptive bearing fault diagnosis for target domain class missing based on conditional Wasserstein generative adversarial network with gradient penalty

Abstract In the field of bearing fault diagnosis, domain adaptive methods usually assume that the source and target domains have the same class distribution, but this assumption does not validate in practical application scenarios. The reason is that the target domain training set cannot contain all fault classes, which makes it impossible to accurately measure the difference between the source and target domains during feature alignment, which in turn leads to the performance degradation of the diagnostic model. To address this problem, this paper proposes a domain adaptive bearing fault diagnosis for target domain class missing based on conditional Wasserstein generative adversarial network with gradient penalty (CWGAN-GP), which is divided into three stages. First, an enhanced CWGAN-GP is constructed by incorporating the class constraint mechanism, which enhances the diversity and authenticity of the generated data, thus solving the problem of missing classes in the target domain. Second, the domain-invariant features learned in the source domain are transferred into the target domain, and the feature distributions in the source and target domains are aligned by introducing multi-kernel maximum mean difference (MK-MMD) in the different fully-connected layers in order to reduce the distributional difference. Finally, a joint loss function is designed to update the network parameters via back propagation to further improve the diagnostic performance of the model. The proposed method is validated on CWRU and MFS datasets with an average accuracy of 99.46% and 99%, respectively, proving its effectiveness in the case of missing classes in the target domain.