Domain generalization methods can effectively identify mechanical faults under unseen working conditions. However, most of them require data from multiple source domains for model training. Nevertheless, the collection of monitoring data of each health state under different working conditions is unpractical for wheel-set bearings. Aiming at the difficulty of obtaining multi-source domain data, a single-source domain generalization model based on flow model and contrastive learning is proposed for fault diagnosis of wheel-set bearings under various working conditions. The proposed model employs flow model as a domain generation module to generate samples in an extended domain. Then, domain-invariant features are extracted from the source domain and the extended domain. The diversity of the generated samples and the effectiveness of the domain-invariant features are guaranteed by a strategy of adversarial contrastive learning. Finally, single-source domain generalization fault diagnosis experiments carried out on a wheel-set bearing dataset verify the good performance of the proposed method over the traditional domain generalization methods.

single domain generalization, fault diagnosis, wheel-set bearing, flow model, contrast learning