| Abstrakt: |
Unsupervised domain adaptation (UDA) has been widely exploited for machinery fault diagnosis (MFD). However, existing UDA approaches always require direct access and sharing to the labeled source domain, raising privacy concerns. In this article, a practical and challenging scenario, source-free UDA (SFUDA), is considered for privacy-preserving MFD. In SFUDA, only a pretrained source model is provided for the unlabeled target domain, and the source data are inaccessible during adaptation. A novel SFUDA approach, namely source-free cluster adaptation (SF-CA), is proposed, which consists of source domain generalization (SDG) and target model adaptation (TMA). SDG aims to obtain a well-generalized source model for TMA. Specifically, the adaptive R-drop is proposed for SDG, where an entropy-aware weighted consistency training strategy is introduced to regularize dropout. TMA enforces cluster assumption for adaptation, where the structure exploration regularizations are proposed to learn structural information. Besides, selective self-training is introduced to alleviate model collapse during TMA. Finally, discriminative and tight-clustered target features can be obtained for SFUDA. Extensive experiments are conducted on three public datasets and one practical dataset. The experimental results show the effectiveness of SF-CA for privacy-preserving MFD and the feasibility of cluster assumption enforcement for SFUDA. |
