Abstract:Data-driven intelligent methods have achieved notable performance in the field of bearing fault diagnosis under stationary conditions. However, in some actual scenarios such as high-speed trains accelerate or decelerate, it is hard to implement traditional intelligent fault diagnosis methods due to the lack of fault data and the nonstationary operating conditions. Although domain generalization-based methods have been proposed for fault diagnosis without target fault data, most researches focus on stationary or segmented stationary working conditions, ignoring the fact that working condition is varying continuously, which limits their successes in practical application. To solve the intractable problem, this article proposes an information-induced feature decomposition and augmentation framework (IIFDA) to generalize diagnosis knowledge from stationary working conditions to unseen non-stationary working conditions. In IIFDA, an information-induced feature learning network (IIFLN) is proposed to infer information-related distributions, and its rationale is also analyzed. Furthermore, an augmented feature synthesis (AFS) method with two feature augmentation techniques, extrapolation (EP) and gradient confusion (GC), is proposed to increase the diversity of training features and help network to refine fault-related features for better generalization. Finally, two bearing case studies are conducted under multiple non-stationary working conditions, which indicate the IIFDA is superior to widely used methods.

An Information-Induced Fault Diagnosis Framework Generalizing from Stationary to Unknown Nonstationary Working Conditions