SATT: Partial Transfer Machinery Fault Diagnosis Based on Subdomain Adaptation and Two-Stage Training

Abstract

In recent years, unsupervised transfer learning based on deep domain adaptation techniques has successfully solved the domain shift problem in machinery fault diagnosis, with a basic assumption that target and source domains share the same label space. However, the label space of the target domain is usually a subset of the source domain in real industrial scenarios, which is beyond the capability of the existing methods. This paper proposes a new partial transfer diagnosis method based on subdomain adaptation and two-stage training (SATT) to solve this problem. In this non-adversarial method, subdomain adaptation is introduced to align the distribution of relevant subdomains in the source and target domains by minimizing the local maximum mean discrepancy between the two domains. Subdomain adaptation can effectively reduce the negative impact of source outlier classes. Additionally, SATT integrates two networks and uses the entropy minimization principle to provide more accurate target pseudo labels. Finally, a two-stage training strategy is designed to remove most of the source outlier samples. Thus, the influence of source outlier classes is further reduced. Experimental results on two diagnosis datasets show that SATT achieves more satisfactory performance than six representative comparison methods.