Cross domain fault diagnosis method based on discriminative geometric distribution alignment

Accurate fault diagnosis of rolling bearings under varying operating conditions remains challenging due to distribution shifts in vibration data. While feature-based transfer learning methods mitigate this issue, many suffer from limitations: over-reliance on maximum mean discrepancy with non-characteristic kernels, which may discard critical distribution information, and the assumption of a single shared transformation matrix for domain alignment, which fails under significant distribution shifts. To address these shortcomings, a discriminative geometric distribution alignment (DGDA) model is proposed for cross-domain fault diagnosis. DGDA simultaneously optimizes two distinct transformation matrices for source and target domains. It integrates the dynamic maximum mean discrepancy, maximum covariance discrepancy to establish a refined distribution discrepancy metric, effectively capturing complex data differences. The model further preserves geometric structures by enhancing inter-class separation and intra-class compactness in source domain while maximizing feature variance in target domain. Finally, based on the new feature representations, a pattern recognition method is used to diagnose the target domain samples. The fault diagnosis experiments under variable operating conditions and cross-machine verify that the proposed model can improve the domain adaptation between source and target domains, enhance the transfer ability of source domain diagnosis knowledge, and realize cross-domain fault diagnosis under different data distributions more accurately.