A Causality-Inspired Capsule Network for Domain Generalization in Cross-Domain Fault Diagnosis of Rolling Bearing

Cross-domain fault diagnosis (CDFD) leveraging domain generalization (DG) has attracted increasing attention due to its potential to generalize diagnostic knowledge to unseen data domains. However, the majority of current DG methods primarily model statistical dependencies, which limits their generalization performance in unobserved domains. To address this limitation, causal learning offers a natural advantage in revealing the causal linkages within diagnostic knowledge, providing a novel perspective for diagnostic guidance. Inspired by this strength, this study proposes a causality-inspired capsule network (CICN) for cross-domain fault diagnosis in the bearing field. The CICN employs a capsule network to extract features that encapsulate rich fault-related information. During the dynamic routing process, a causal gated unit is introduced to emphasize the potential causal diagnostic knowledge embedded in multi-source domain fault data, enabling the model to leverage richer causal information. Furthermore, the method disentangles causal features from noncausal features and enhances the robustness of causal invariant features through causal intervention, improving the model’s performance against domain interference. The effectiveness of the CICN model is validated through cross-domain fault diagnosis experiments on five datasets. The results demonstrate the superior generalization capability of the proposed approach across diverse working conditions and different machine settings.