Joint domain transfer elasticity metric network for cross-domain small sample fault diagnosis

Inadequate target domain samples are successfully addressed by transfer learning, but it suffers when there are also inadequate source domain samples. Meta-learning offers a novel solution to the small-sample problem by enhancing model generalization through multi-task learning. However, current meta-transfer research has not adequately considered the domain distribution differences resulting from changes in operating conditions and the problem of insufficient source domain samples, while also facing the challenge of class confusion caused by feature complexity and fuzzy class boundaries. To overcome these difficulties, this study proposes a joint domain transfer elasticity metric network (JDTEMN) to solve the cross-domain small sample fault diagnosis problem. First, a joint cross-domain transfer structure is introduced, utilizing maximum mean square difference (MMSD) with domain discriminators to match source and target domain features. This approach leverages sample mean and variance for accurate feature matching and incorporates adversarial training to capture domain-specific features, enabling more effective transfer. Second, the metric-based meta learning framework is integrated with the joint cross-domain transfer structure to address the small sample problem. To resolve class confusion in the metric network, an elasticity factor is introduced to reduce intraclass spacing and increase interclass spacing, enabling effective cross-domain small-sample fault diagnosis. To enhance the network’s learning of fault features, the CBAM attention mechanism is incorporated into JDTEMN, further improving diagnostic accuracy. Finally, the effectiveness of JDTEMN is validated through experiments on public and private bearing datasets, demonstrating superior diagnostic performance compared to other methods.