Domain-invariant feature learning based on a curvature-sensitive regulation strategy for bearing fault diagnosis under variable working conditions

Abstract In recent years, substantial progress has been achieved in addressing cross-domain fault diagnosis problems through knowledge transfer strategies. However, traditional studies usually assume that target data are provided to facilitate domain adaptation throughout the training process, which is often infeasible in practical engineering applications, where fault data are often scarce. Therefore, extracting general and effective diagnostic leveraged knowledge from the source domain and applying it to unknown target tasks are crucial for intelligent cross-domain fault detection. To this end, this paper introduces a curvature-sensitive regulation strategy (CSRS) for cross-domain generalization of domain-invariant features in bearing fault diagnosis. To ensure high diagnostic accuracy in previously unobserved domains, this study proposes a model that converges to a smooth region with low loss values, allowing for the extraction of domain-invariant features, distinguishing features from source domains that may differ, which can be maintained in new target domains without requiring the presence of fault data. The effectiveness, superiority, and robustness of the CSRS method have been confirmed in generalization tasks across various fault diagnosis domains.