A domain adaptation bearing fault diagnosis method with strong noise resistance

To address the problem of the weak anti-noise capability of domain adaptation methods in transfer learning tasks for bearing fault diagnosis, this study proposes a domain adaptation-based bearing fault diagnosis method with strong anti-noise ability. In this method, a one-dimensional multi-scale efficient channel attention residual network (1D-MS-ECA-ResNet) is first designed. The 1D-MS-ECA-ResNet features a multi-scale structure and an efficient channel attention residual structure. The multi-scale structure employs large-kernel convolutions, which can cover a longer time window of the input signal, enabling smoother feature extraction of the input signal and reducing the interference of high-frequency noise. The channel attention residual structure can effectively screen out the bearing fault features from the features extracted by the multi-scale structure, further minimizing the impact of noise on the process of fault feature extraction. Subsequently, an improved discriminative joint probability maximum mean discrepancy (IDJP-MMD) mechanism is proposed. In the IDJP-MMD mechanism, to mitigate the interference of noise on domain adaptation, the discriminative joint probability maximum mean discrepancy (DJP-MMD) and the deep domain adaptation correlation alignment (CORAL) used for measuring second-order statistics are combined through a dynamic balance parameter to form a new distribution difference indicator. In addition, adversarial domain adaptation and IDJP-MMD are simultaneously applied to the adaptation process of features from the two domains, so as to achieve the alignment of features from the two domains from multiple perspectives. Finally, cross-condition transfer learning experiments are conducted using two different datasets under various noise backgrounds. The results indicate that, even in the presence of noise interference, the method proposed in this paper can effectively enable a good adaptation of features from the two domains, which holds potential value for practical engineering applications.