Open Set Bearing Defect Recognition Based on Selective Momentum Contrastive Dual Adversarial Learning

The unpredictability of machine failure modes and locations in practical engineering applications can severely disrupt the diagnostic task of data-driven closed-set defect recognition methods. To address this issue, this paper studies the defects transfer diagnosis of rolling bearings under open-set scenario. We propose a Momentum Contrastive based Dual Adversarial (MCDA) domain adaptive method based on global difference degree comparison, which performs unsupervised clustering and separation of preferred negative samples based on difference measurement in all domains based on contrastive learning, and then in the source and target domains by Adversarial domain classifiers and feature extractors enable domain adaptation, where encoders and feature extractors share weights for a second adversarial learning. Clustering is accomplished with private classes preferred separation in the first step of contrastive learning, and the alignment of shared classes is achieved in the second step by domain adaptation through semi-supervised training, whereby the target private classes are separated. Our experiments on the Paderborn Dataset demonstrate the effectiveness of this method.