In the field of domain generalization for fault diagnosis, the majority of approaches concentrate on extracting domain-invariant features from multi-source domains. However, collecting samples from multi-source domains is extremely difficult, and the data typically originate from a single-source domain. To tackle the issue of inadequate generalization capability in unknown target domains when trained on only a single source domain, an improved prototype-guided causal disentanglement domain generalization network (ICDDG) is proposed for mechanical fault diagnosis. This network combines feature mean, similarity, and triplet loss to construct an improved prototype-based triplet loss function, which reduces the influence of outlier samples and achieves more effective prototype learning. The improved triplet loss function effectively guides the causal disentanglement network to separate causal features from non-causal features better, enhancing the model's adaptability and robustness when encountering unseen domains. Diagnostic experiments performed using two bearing datasets substantiate the efficacy of the ICDDG method.

causal disentanglement;single-source domain generalization;bearing;fault diagnosis;triplet loss