Dual-Weight Consistency-Induced Partial Domain Adaptation Network for Intelligent Fault Diagnosis of Machinery

Abstract

Domain adaptation-based methods have been broadly developed for cross-domain fault diagnosis of machinery, in which the target and source label spaces are identical. However, a more challenging and practical diagnostic scenario in which the source label space subsumes target one is remaining to be solved. To address this issue, an innovative two-stage dual-weight consistency-induced partial domain adaptation (DCPDA) network, consisting of two feature extractors, a classifier, a domain discriminator, a dual-weight consistency-induced reweighting (DCR) module, and a Wasserstein distance based-domain adaptation (WDA) module, is proposed. In the pre-trained stage, a well-trained model on the source domain is acquired. In the fine-tuned stage, the DCR, which conducts a dual-weight consistency-induced weighting strategy, obtains a bi-level compound weight from both class-level and sample-level weights. Embedded with the bi-level weight, the WDA is encouraged to transfer the source knowledge with mutual agreement from double-level weights across domains. Accordingly, the DCPDA is capable of alleviating negative transfer dramatically and mapping the source diagnostic knowledge into the target domain selectively. Extensive experiments indicate the validity and advantage of the proposed DCPDA.