A bearing fault diagnosis method based on contrastive disentanglement for single-source domain generalization

Abstract Changes in operating conditions cause distribution shifts in monitoring data. In practical engineering scenarios, collecting training data that cover all operating conditions is often infeasible, which limits fault diagnosis methods that rely on multi-source data or prior target-domain information. Under single-condition training, suppressing condition-related non-causal features and extracting causal representations shared across conditions remain key challenges for cross-condition diagnosis. To address this problem, a contrastive disentanglement single-source domain generalization network is proposed. First, a contrastive disentanglement framework is developed based on causal disentanglement theory and the bootstrap your own latent contrastive learning network. To achieve causal feature extraction, a dynamic weight redundancy reduction loss is proposed based on the Hilbert–Schmidt independence criterion. Second, an adaptive threshold-weighted nuclear norm regularization strategy is further proposed to constrain feature rank and promote shared causal representations across domains. Finally, a prototype Gaussian triplet loss function is designed to improve domain invariance by optimizing the prototype distribution in feature space. Experiments on three bearing datasets demonstrate that the proposed method consistently outperforms seven state-of-the-art approaches. The average diagnostic accuracies are 97.99%, 98.82%, and 83.33%, respectively. These results exceed those of the best competing methods by 3.80%, 0.82%, and 4.49%. In addition, the average standard deviations are as low as 0.04%, 0.15%, and 0.25%, indicating superior stability.