PC3Net: A prior-causal contrast-collaboration network for single domain generalization fault diagnosis

Most current domain generalization diagnosis methods improve the diagnostic performance of models in unknown target domains by training with numerous source domains. However, it is challenging to obtain comprehensive monitoring data from industrial sites. Therefore, a robust fault diagnosis model capable of generalizing from a single source domain to multiple unseen target domains holds greater engineering application value. The paper proposes a prior-causal contrast-collaboration network (PC3Net) for single domain generalization (SDG). To elucidate domain-general causal mechanisms, a causal feature extractor guided by a dual causality loss function under natural causal intervention is constructed to purify fault-related causal features. A prior-causal contrastive loss function is innovatively proposed, based on low-dimensional static prior features and high-dimensional dynamic causal features, to enhance the consistency of multi-dimensional features, assisting the causal feature extractor in aligning with dynamic response-aware prior knowledge while focusing on high-dimensional complex relationships. A dual-feature collaborative cross-entropy loss function is constructed to enhance the decision-making ability of the diagnosis model. Multiple datasets are used to construct comparative, ablation, and model interpretability analysis experiments based on Explainable Artificial Intelligence technologies. Results demonstrate that the proposed method outperforms state-of-the-art approaches in both generalization and interpretability for SDG diagnosis tasks.