Multi-source domain self-supervised enhanced transfer fault diagnosis approach with source sample refinement strategy

Existing research in multi-source transfer learning often overlooks the differences among source domain samples and fails to effectively utilize unknown target domain information. To address these limitations, this paper proposes a multi-source domain self-supervised enhanced transfer fault diagnosis method with source sample refinement strategy. Firstly, a joint training mechanism is introduced to train all source predictors, leveraging knowledge from all source domains. Subsequently, a source sample refinement strategy is devised, utilizing a category classifier in conjunction with predefined thresholds to filter out source samples highly similar to the target samples, thereby obtaining weights for each source domain. Then, a multi-level distribution matching mechanism is designed to reduce the distribution discrepancy between refined source samples and target samples, followed by the development of a self-supervised training mechanism to further enhance the cross-domain performance of source predictors. Finally, integrating the weights from each source domain, the trained source predictors are combined to derive the target predictor, yielding the final diagnostic results. Experimental results indicate that the proposed method can achieve better target alignment by leveraging fault information from the unknown target domain, effectively addressing cross-domain diagnosis challenges.