Dual disentanglement domain generalization method for rotating Machinery fault diagnosis

• Dual disentanglement networks to extract class-relevant domain invariant features. • Dual contrastive module decoupling domain specific and domain invariant features. • The adversarial mask module decouples class-relevant and class-irrelevant features. • Learnable domain-aware and class-aware masks ensure feature complementarity. The objective of domain generalization fault diagnosis is to develop a robust model that can generalize to unseen domains. This makes it a highly ambitious and challenging task. However, most current methods rely on domain labels to extract domain-invariant features and do not consider the negative impact of the presence of class-irrelevant features in domain-invariant features on generalization. Therefore, this paper proposes a dual disentanglement domain generalization method for rotating machinery fault diagnosis that does not depend on domain labels. Based on the analysis of the potential features between domains and class labels, a dual contrastive disentanglement module and an adversarial mask disentanglement module are proposed to disentangle the domain-invariant and class-relevant features, respectively. Specifically, in the dual contrastive disentanglement module, the concept of contrasting is employed to train the network shallow features of the source data and the style-enhanced data to produce domain-aware mask decoupled domain-specific and domain-invariant representations. The adversarial mask disentanglement module uses an adversarial classifier to update the class-aware mask and further accurately separate class-relevant and class-irrelevant features. Concurrently, the KLD loss is devised to guarantee that the class-relevant features encompass sufficient labeling information. Finally, the efficacy of the method is substantiated by comprehensive experimental findings on both public and private datasets. The code will be available at: https://github.com/GuoweiaaZhang/DDDG .