SDCGAN: A CycleGAN-based single-domain generalization method for mechanical fault diagnosis

In recent years, fault diagnosis based on domain generalization has attracted increasing attention as an effective approach to address the challenge of domain shift. most existing approaches depend on learning domain-invariant representations from multiple source domains, limiting their practical application in fault diagnosis. To address this issue, this paper introduces a single-domain generalization method for mechanical fault diagnosis, the Single-Domain Cycle Generative Adversarial Network (SDCGAN). A CycleGAN-based domain generation module is introduced to produce extended domains that exhibit substantial divergence from the source domain, enhancing the model's generalization capability. The diagnostic task module subsequently extracts domain-invariant features from both the source and extended domains. Furthermore, an adversarial contrastive training strategy is employed to learn generalized features robust to unknown domain shifts. Comprehensive experiments on two mechanical datasets verify the effectiveness of the proposed method, while ablation studies validate the contributions of its components, highlighting its potential for real-world applications.