Causal Meta-Learning Framework: A Co-Evolutionary Approach to Disentanglement and Generalization for Fault Diagnosis of Rotating Machinery

The generalization of intelligent fault diagnosis models under unseen operating conditions remains a significant challenge due to the complex and variable environments of rotating machinery. While recent domain generalization (DG) methods have mitigated performance degradation from distribution shifts, they often rely on generic architectures and indirect statistical constraints. This makes it difficult to disentangle causal fault features from domain-specific spurious correlations; furthermore, these methods typically lack a holistic, generalization-oriented optimization mechanism. To address these limitations, this article proposes a causal meta-learning framework (CMLF) that systematically enhances generalization through synergistic designs in its architecture, constraints, and optimization. At its core, CMLF employs a heterogeneous causal disentanglement architecture (HCDA) to explicitly separate stable causal features from spurious domain-related features. This is reinforced by a multitask collaborative constraint system designed to effectively block spurious association paths and strengthen the domain invariance of causal representations. The entire framework is embedded within a meta-learning paradigm that simulates unseen conditions, thereby synergistically enhancing both disentanglement and generalization capabilities. Extensive experiments on component- and system-level datasets demonstrate that CMLF significantly outperforms state-of-the-art methods, achieving average cross-domain accuracies of 95.30% and 94.82%, respectively. These results confirm its excellent and robust diagnostic performance under unseen conditions.