A centroid contrastive multi-source domain adaptation method for fault diagnosis with category shift

Abstract

Multi-source domain adaptation for fault diagnosis aims to transfer knowledge from multiple source domains to the target domain to enhance the reliability and safety of equipment. Current multi-source domain adaptation methods primarily focus on addressing domain shift while overlooking category shift across different domains, which leads to diagnosis performance degradation in real-world scenarios. To tackle this issue, a centroid contrastive multi-source domain adaptation (CCMDA) method is proposed for fault diagnosis with category shift. The model consists of a common feature extraction module and a multi-source domain adaptation module. The feature extraction module extracts feature from multiple sources and the target domain. The multi-source domain adaptation module mitigates domain shift through adversarial training and addresses category shift using centroid contrastive loss. Experimental results on both label-consistent and label-inconsistent multi-source transfer tasks demonstrate the effectiveness of the proposed model.