Transferable mapping shift network for unsupervised domain adaptation using in vibration signal fault diagnosis under variable conditions

Unsupervised domain adaptation as one of the popular techniques of deep transfer learning used to tackle the model performance suffers the domain gap between the source and target data, and it is recently developed for bearing fault diagnosis under variable working conditions. Existing methods focus on learning the domain-invariant features but seldom considering whether the representations are transferable for domain adaptation. To address the mentioned problem, a novel adversarial-based two-branch transferable mapping shift network (TMSN) is proposed by introducing two-granularity feature fusion and transferable feature adaptive selection methods. Specifically, first, a two-branch ResNet34 baseline and a Transformer-based concatenation module are integrated to extract deep features of the raw vibration signals. Second, a multi-head attention-based transferable mapping shift (TMS) module is employed to adaptatively enhance the transferable features and weaken the non-transferable ones at different key layers of feature extractor. Furthermore, a domain adaptor is used to measure the feature discrepancies of different domains. Experiments are conducted on public dataset, the results present that TMSN obtains the best performance compare with other state-of-the-art approaches. The visualizations of distribution discrepancy and attention weights of TMS modules imply that TMSN can adaptively select the transferable features and effectively align the source and target domains. The proposed TMSN is effective for the bearing diagnosis tasks under variable conditions.