An Unsupervised Domain Adaptation Method for Intelligent Bearing Fault Diagnosis Based on Signal Reconstruction by Cycle-Consistent Adversarial Learning

Conventional deep learning trained with labeled data has made remarkable achievements in the field of mechanical fault diagnosis. Nevertheless, in practical industrial applications, it is hard to obtain enough, identically distributed, and well-labeled monitoring data for training a well-generalized network. To relax these restrictions on training data, this article proposes a new domain adaptation fault diagnosis method that fuses cycle-consistent adversarial (CCA) learning and multiple kernel variant of the maximum mean discrepancy (MK-MMD) distance measure algorithm. The CCA learning intends to align distributions in the generation space. MK-MMD metric in domain-shared module closes across-domain distances by learning invariant features in high-level embedded spaces. To facilitate the reconstructed samples with smaller intraclass distance and greater interclass distance, the proposed method adopts a new training strategy, that is, the CCA module is no longer trained independently, but trained jointly with the domain-shared module and classifier. To reduce consumption in making datasets, bearing vibration signals collected by acceleration sensors are directly used after being standardized and segmented. It is demonstrated that the fusion improves the aggregation of reconstructed samples and promotes positive transfer across domains. The comparison with other transfer learning methods verifies the effectiveness of the proposed method and shows its superiority in diagnostic performance.