Adaptive mask flow adversarial network for single-source domain generalization fault diagnosis of train bearings

Fault diagnosis of train bearings is crucial for railway safety, yet models trained on a single operating condition often experience significant performance degradation when subjected to variations in speed or load. This paper proposes a new single-source domain generalization (SDG) model named adaptive mask flow adversarial network (AMFAN), aiming to enhance generalization capability by effective cross-domain simulation based on learnable perturbations in feature space. An adaptive mask mechanism is designed to determine domain-sensitive feature elements. And a feature perturbation strategy is conducted to the determined feature elements via a pre-trained flow model. Experiments on two train bearing datasets verify the superior performance over state-of-the-art methods. The results prove that the controlled feature expansion provides a viable and robust pathway for SDG, showing strong potential for real-world train bearing fault diagnosis when confronting unknown operating conditions.