Decoupled feature learning framework with class-guided adversarial alignment for rotating machinery fault diagnosis

Existing multi-source domain methods tend to directly map raw data into a shared subspace, while neglecting explicit decoupling between fault-related and non-related information. Moreover, these methods primarily enforce domain-level alignment without introducing class-level constraints, which leads to blurred inter-class boundaries and unexcellent discrimination under domain shift. To tackle these problems, this paper proposes a decoupled feature learning framework with class-guided adversarial alignment (DFLFCAA). Firstly, a decoupled feature generator with channel-wise adaptive thresholds is designed to decompose the learned features into fault-related and non-related components. Only the former are retained for subsequent diagnostic tasks, thereby eliminating domain-specific interference at the feature level. Secondly, a class prototype contrastive alignment strategy is introduced, which constructs cross-domain class prototypes and enforces three constraints to achieve fine-grained class-level structural alignment in the shared subspace. Thirdly, a task-aware domain discriminator is developed by incorporating class prediction probabilities into the discriminator input, which enables category-conditioned domain judgment that prevents the confusion between class variance and domain variance. Extensive evaluations on multiple datasets confirm that DFLFCAA achieves superior fault identification accuracy and robustness under scarce-label scenarios.