Stochastic Embedding Domain Generalization Network for Rotating Machinery Fault Diagnosis Under Unseen Operating Conditions

Domain generalization-based fault diagnosis (DGFD) methods have been extensively explored in cross-domain fault diagnosis under various operating conditions in recent times. Nevertheless, these methods adhere to a common premise that the fault modes across each available source domain remain consistent. The label inconsistent problem arises when the model extracts domain-invariant features from multiple source domains. That is to say, the fault modes between source domains are inconsistent, resulting in overfitting to scarce fault modes during model training. Aiming at this problem, this study presents a stochastic embedding domain generalization network (SEDGN) for rotating machinery fault diagnosis, particularly in scenarios, where inconsistent source fault modes exist across multiple source domains. First, a stochastic embedding layer is designed to mitigate the overfitting to scarce fault modes, in which the weights of the fault identifier for each fault mode are modeled by Gaussian distributions and will be optimized during model training. Second, a ground-truth label-guided correlation alignment is further introduced for shared fault modes across multiple source domains, enhancing the domain-invariant fault feature extraction between shared fault modes. Finally, a gearbox fault dataset containing bearing and gear faults is utilized to simulate domain generalization tasks under label inconsistent problems, and the effectiveness of the proposed SEDGN methods is further validated.