A novel domain generalization network integrating invariance and cohesiveness for rolling bearing fault diagnosis

Currently, fault diagnosis based on domain generalization (DG) for unseen target domains has become a prominent research topic. Despite significant progress, several challenges remain, including domain-dependent features in the extracted features and the overlap of feature clusters across different classes. To address these challenges, we propose a novel DG network that integrates invariance and cohesiveness. First, based on the information bottleneck, we present a domain-invariant feature loss that compresses redundant features and eliminates domain-dependent ones. Second, by employing feature segmentation and maximum mean discrepancy, we propose a new feature cohesion regularization method that enhances the model’s clustering performance. Finally, the feature invariance and cohesion losses are integrated into a unified framework. Experiments conducted on two fault datasets demonstrate that the method exhibits strong generalization diagnostic capability and high diagnostic accuracy, and can be applied to fault diagnosis of rolling bearings under unseen operating conditions.