Open-Set Domain Adaptation via Feature Clustering and Separation for Fault Diagnosis

In the realm of fault diagnosis, a challenge arises when the target domain (TD) introduces new fault categories not present in the source domain (SD). This challenge is referred to as the open-set domain adaptation (OSDA) fault diagnosis issue. This study proposed a novel OSDA method using feature clustering and separation (FCS-OSDA) to address this problem. Entropy minimization-only (EMO) and diversity maximization (DM) were employed as the basic elements for the FCS-OSDA. These components were efficiently implemented through standard stochastic gradient descent, eliminating the need for adversarial learning. Moreover, a novel memory module based on momentum update was designed to assess the similarity of neighboring features with the aim of achieving well-clustered features for known classes. As a supplement, a pseudo-decision boundary was established by incorporating a joint-entropy loss, which separated the features of known and unknown classes. Extensive experiments on three rolling bearing datasets validated the effectiveness of the FCS-OSDA in addressing the OSDA fault diagnosis issue. Various analysis results demonstrated the superiority of the FCS-OSDA over its state-of-the-art competitors.