A multi-target domain adaptive method for intelligent transfer fault diagnosis

Abstract

Existing domain adaptive fault diagnosis methods hardly consider single-source-multi-target scenarios, but the single-source-multi-target model can reduce the use of labeled data and enhance information sharing, with strong practicability. However, applying single-target models directly to multi-target domain adaptation is likely to be suboptimal since conventional methods cannot directly connect the shared knowledge among multiple domains. In this work, we propose a multi-target domain adaptation (MTDA) method for fault diagnosis of rotating mechanical components, which improves the transfer accuracy by using the domain features of multiple target domains. Firstly, a transferable hierarchical sparse autoencoder (THSAE) method is proposed for extracting domain-invariant features from single-source and single-target domains, which combines a hierarchical sparse strategy, maximum mean discrepancy, as well as class aggregation and separation strategy. Secondly, a multi-branch transfer strategy is proposed to fuse the three THSAE extracted features in a domain fusion to improve the model’s generalization ability. The results of two sets of experimental cases show that the proposed method not only demonstrates reliable and stable diagnostic performance but also can realize the diagnosis of multiple target domains with one model, which has certain practicability.