Gradient aligned domain generalization with a mutual teaching teacher-student network for intelligent fault diagnosis

Domain generalization for intelligent fault diagnosis handles multiple source domains and generalizes learned knowledge to unseen target domains. Recent efforts actively advocate for learning domain invariances that retain only domain-invariant features and reduce domain-specific ones. However, most of their invariance criteria might not be effective, compared with the simple empirical risk minimization. Besides, pursuing a similar training procedure has shown the potential to promote domain invariances, but it involves gradient alignments that are quite hard to manipulate. To address these issues, a Gradient Aligned Domain Generalization method (GADG) with a mutual teaching teacher-student network is proposed. Specifically, domain-invariant features are first thoroughly identified using gradients derived from multiple task-specific classifiers, which also intend to offer structure benefits by taking disentangled features for respective tasks. Then, to capture highly expressive domain invariances in the gradient space, a granular gradient alignment is initiated to emphasize high-order domain consistencies, which collaborate with low-order invariant distributional statistics to guarantee safe gradient alignments. Finally, to satisfy consistent gradient alignments, a mutual teaching teacher-student network is proposed, in which modules are assigned with respective tasks and recursively trained. Comprehensive experiments on various generalization tasks validate the efficacy of GADG.