Domain generalization network for intelligent fault diagnosis based on low-rank Gaussian feature learning

Diagnosis of bearing faults is an indispensable component of health monitoring for rotating machinery. However, distribution shifts induced by operating-condition changes severely degrade diagnostic generalization. Domain generalization (DG) learns exclusively from source-domain data without requiring target-domain participation. To this end, this paper develops a coordinated domain-generalization framework, LGFDG-Net, rather than a collection of isolated, standalone operators. Its multi-level data-augmentation pipeline sequentially applies SegmentMix and instance mixing to expand local temporal structures and global sample distributions. Multi-branch channel-spatial attention extracts fault responses at complementary receptive fields. Kullback–Leibler divergence and selective low-rank regularization are imposed on stochastic latent features, together with a dynamically weighted source-domain adversary. This design is intended to attenuate measurement deviations and redundant latent directions induced by working-condition variations and noise, while retaining transient fault structures. Across eight held-out cross-condition tasks from two public laboratory datasets, evaluated under a source-only protocol with − 10 dB additive Gaussian noise, LGFDG-Net exceeds the strongest evaluated baseline by 12.26 and 15.53 percentage points in average accuracy on CWRU and PU, respectively. The evidence is therefore limited to the studied datasets, operating conditions, and noise model; the ablation results indicate the existence of task relevance. Validation on independent machinery, real acquisition campaigns, and diverse noise types remains necessary before field deployment can be assessed.