Bearing fault diagnosis method based on contrastive learning and domain adaptation under variable working conditions

Most bearing fault diagnosis methods based on transfer learning for variable working conditions typically focus on domain alignment, while neglecting the class information of the samples themselves. This oversight leads to inaccurate class alignment between the source and target domains, thereby reducing diagnostic accuracy. To address this issue, a novel fault diagnosis method based on contrastive learning and domain adaptation network (CDAN) is proposed. Firstly, a deep residual shrinkage network with channel-wise thresholds (DRSN-CWT) is utilized to directly extract features from raw vibration data, thereby maximizing the extraction of relevant features. Subsequently, a feature contrast module, guided by a novel global contrastive loss (GCL), is introduced to quantify the similarity between different extracted feature data distributions, performing contrastive analysis on the extracted features to maximize the distance between samples of different fault classes and minimize the distance between samples of the same fault class. Concurrently, an adversarial domain adaptation module is utilized to learn the discriminative features shared between domains, aligning the data distributions of the source and target domains. Furthermore, an adaptive factor is designed to dynamically balance the relative importance between domain alignment and classification performance, mitigating the adverse impacts caused by overly large or small loss terms. Experimental results on the CWRU and PU bearing datasets validate the effectiveness and superiority of the proposed method, achieving average diagnostic accuracies of 99.64% and 80.25%, respectively.