Deep Causal Disentanglement Network With Domain Generalization for Cross-Machine Bearing Fault Diagnosis

Domain generalization-based fault diagnosis (DGFDs) has gained considerable attention in bearing fault diagnosis due to its ability to extract feature-invariant information from diverse source domains, without requiring direct access to target domain data. However, many existing DGFD approaches primarily rely on statistical models to capture the relationship between time-series data and labels. This often leads to the learning of entangled features, as these models lack prior knowledge to differentiate between task-relevant and task-irrelevant information. To address this limitation, this article introduces the deep causal disentanglement network (DCDN), a novel approach tailored for cross-machine bearing fault diagnosis. In this framework, fault data collected from multiple source domains is decomposed into causal factors related to fault representation and non-causal factors associated with domain-specific information, using a structural causal model (SCM). This process effectively reconstructs the data generation pathway. By optimizing causal aggregation loss and maximizing information entropy loss, DCDN can distinguish between causal and non-causal features from both direct and indirect perspectives. Furthermore, a contrastive estimation loss is minimized to ensure that the extracted causal features retain most of the essential information from the original dataset. Additionally, a redundancy reduction loss is employed to minimize correlations among the dimensions of the causal vector, further reducing the entanglement between causal and non-causal factors. The effectiveness and superiority of the proposed model are demonstrated across five cross-machine bearing fault datasets. Experimental results show that, compared to other state-of-the-art (SOAT) methods, DCDN achieves superior performance in both estimation accuracy and robustness.