Learn Generalized Features Via Multi-Source Domain Adaptation: Intelligent Diagnosis under Variable/Constant Machine Conditions

A primary goal of fault diagnosis is to build generalizable models for flexible industrial scenes. However, most literature assumed that the training and testing data are collected from the same distributions, which poses obstacles in cross-domain diagnosis. This paper proposes a multi-source domain adaptation (MSDA) strategy in response to diversiform working conditions, in which the labeled testing data are difficult to obtain. The regularization term expressed by multi-order moment matching is adopted to extract transferable knowledge from multiple source domains. An adversarial strategy, which takes the maximization and minimization of two independent classifiers' discrepancy as the alternate optimization objective, is introduced to dynamically align moments of feature distributions between all domain pairs. Three datasets are carried out to prove the robustness of the proposed method. The results of comparative experiments and ablation study demonstrate that the approach can learn features from multiple specific domains and possess strong generalization ability under diverse constant or any unknown variable conditions with unlabeled target samples.