A multi-order moment matching-based unsupervised domain adaptation with application to cross-working condition fault diagnosis of rolling bearings

Unsupervised domain adaptation-based transfer learning (TL) has been widely used in rolling bearing fault diagnosis to overcome the problem of limited and non-identically distributed labeled data. Discrepancy-based alignment is a popular domain adaptation method in TL. However, due to the inability to completely eliminate domain drift, the classifier learned from the source domain may easily misclassify some target domain samples that are scattered near the decision edge. In this work, a multi-order moment matching-based domain adaptation is proposed to address the issue. Low- and high-order moment matching is simultaneously applied to describe the complex non-Gaussian distributions in more detail and realize coarse- and fine-grained hybrid domain alignment. Furthermore, a discriminative clustering approach is employed to extract domain-invariant features of inter-class discrimination and intra-class compactness, which effectively reduces the negative transfer caused by hard-aligned target samples. The application of the proposed model to the experimental dataset demonstrates that the model can significantly improve the diagnosis accuracy of rolling bearing faults in cross-working conditions. This study can be of assistance to engineers in promptly identifying and addressing rolling bearing faults, ultimately enhancing the reliability and safety of equipment.