CSD-CGCF: A Causal Semantic Disentanglement and Cross-Graph Constraints Framework for Category Overlap Suppression in Zero-Shot Compound Fault Decoupling

The prevalence of compound faults in the health monitoring of complex industrial equipment poses significant challenges, particularly due to the scarcity of labeled samples and the high cost of annotation. This has stimulated considerable research interest in Zero-Shot Learning for compound fault decoupling. However, existing methods often overlook the impact of feature space overlap between compound faults and their constituent individual faults, which severely degrades decoupling accuracy. Accurately decoupling compound faults under this condition remains a significant challenge. To address this issue, we proposes a compound fault decoupling framework based on causal semantic disentanglement and cross-graph constraints. First, a cross-graph constraint mechanism is introduced, integrating the causal feature graph and the fault semantic graph. This facilitates the fusion of causal contributions and semantic structures within the feature space. Furthermore, we design a causal decoupling classification network based on contrastive learning. This network leverages counterfactual reasoning to generate decoupled fault features and employs a hierarchical prototype contrastive loss to establish a metric space with clear inter-class boundaries. Finally, a causal-aware classification network is built using contrastive similarity functions to achieve high-accuracy compound fault decoupling in zero-shot scenarios. Experiments on the motor bearing fault dataset demonstrate that the proposed model outperforms all baseline models in zero-shot compound fault decoupling, with an F1 score of 95.08% ± 0.56. This significantly advances the engineering application of zero-shot compound fault decoupling methods.