Source-free Unsupervised Domain Adaptation for Privacy-Preserving Intelligent Fault Diagnosis

Rolling bearing is of vital significance in industrial applications and Intelligent fault diagnosis (IFD) have been widely exploited in this field. However, cross-machine variations hinder the model performance across different machines in varying working conditions. To solve this issue, many unsupervised domain adaptation (UDA) approaches have been proposed recently, requiring direct access to the fully labeled source domain. However, privacy concerns are raised in traditional UDA scenario, since raw signals contain various private information. To address this issue, the source-free unsupervised domain adaptation (SFUDA) scenario is considered, where only the pre-trained source model is required, instead of the fully labeled source domain. In this paper, a SFUDA approach for intelligent fault diagnosis (IFD) is proposed. It consists of two steps: 1) source model generalization, where virtual adversarial training, R-Drop and label smoothing techniques are adopted to improve the generalization ability of the source model; and 2) target model adaptation, where information maximization is used for cluster assumption and mean teacher training paradigm is utilized to alleviate the catastrophic forgetting phenomenon. The proposed approach is verified on the Case Western Reserve University (CWRU) dataset. Experimental results on the show that the proposed SFUDA approach outperforms several typical UDA approaches, and its effectiveness is verified.