Multi-wise Contrastive Learning with Adversarial Adaptation: Enabling Sample-Efficient Open-Set Fault Diagnosis Across Domains

Unsupervised domain adaptation has made significant strides in cross-domain fault diagnosis. This is predicated on the assumption: sufficient fault samples and target working condition shares the same label space with source one. However, in the realistic scenarios, sample shortage and open-set transfer problem may lead to a significant performance degradation. To address the above challenges, a multi-wise contrastive learning with adversarial adaptation network is designed for few-shot open-set fault diagnosis under variable working conditions in this paper. The core idea is to simultaneously address sample shortage issues and open-set cross-domain fault diagnosis problems. A multi-wise contrastive learning-based prototypical module is designed to extract fault features with limited samples and improve class separation. To tackle significant domain gaps, an adversarial domain adaptation module is designed is designed to align distribution discrepancy between two working conditions. At last, experimental results on few-shot open-set cross-domain bearing fault diagnosis tasks have demonstrated the superiority and effectiveness of the proposed method.