A universal transfer network for machinery fault diagnosis

Abstract

Domain adaptation (DA) methods have achieved promising results in machinery fault diagnosis owing to their ability to mitigate the distribution discrepancy between domains. However, existing fault diagnosis methods based on DA are tailored for a specific setting, and highly rely on prior knowledge about the relationship between the source and target label sets which is usually not available in advance. To broaden the applicability of DA for fault diagnosis, this paper proposes a universal transfer network to handle all types of DA settings, including closed-set DA, partial DA, open-set DA, and open-partial DA. The proposed method utilizes self-supervised learning to uncover the cluster structure of the target domain, and incorporates entropy-based feature alignment to align shared-class samples while separating unknown-class samples. Moreover, an open-set classifier is trained to provide a confidence criterion, which is then used to construct a sample-level uncertainty criterion for identifying unknown-class samples efficiently. The proposed method is evaluated on Office-31 dataset and two fault diagnosis datasets. Our experimental results demonstrate that the proposed method performs better in all DA settings when compared to other methods.