A Novel Contrastive Pre-training-Based Domain Adaptation Method for Fault Diagnosis of Rotating Machines

Data-driven approaches have been widely employed for developing rotating machinery intelligent fault diagnosis systems. Unsupervised domain adaptation methods typically transfer the fault diagnosis model built in the source domain, i.e. common operating condition, to the target domain which is a new operating condition lacking of labeling data. However, in practical applications, only a small proportion data of the source domain data are possibly labeled by experts, leading to the problem of constructing the fault diagnosis model. To address the issue, a contrastive pre-training-based domain adaptation method is proposed. The method extracts discriminative features from source domain data without labels using self-supervised contrastive pre-training, and adversarial domain adaptation is used to obtain domain-invariant features. An experimental platform of the T0 chopper, which is a key device used in the China Spallation Neutron Source, is built and data from the bearing fault injection experiment is collected for verification of the proposed method. Results show that the proposed method has good application prospects in cross-domain diagnosis when only a few data is labeled in the source domain.