A multi-scale feature calibration partial transfer network for rolling bearing fault diagnosis in electric motors

Abstract Domain adaptation approaches have been applied to motor rolling bearing fault detection. Nevertheless, these approaches are constrained by the assumption of label space consistency between source and target domains. In complex industrial application scenarios, it is more common that the target domain label space is a subset of the source domain label space. To address such partial domain adaptation problems, a multi-scale feature calibration partial transfer network is proposed in this paper. This method extracts discriminative features from fault signals through a multi-scale convolutional feature extraction network to improve the accuracy of transfer classification. A domain shift calibration module is designed to explicitly learn cross-domain differences, achieving fine-grained alignment at the feature level. Furthermore, a domain-class joint distribution alignment mechanism is developed. On one hand, this mechanism enhances the transferability of shared classes and suppresses the negative transfer of outlier classes through weighted class alignment. On the other hand, the feature-level and task-level joint distribution alignment mechanisms work synergistically to minimize cross-domain discrepancy. Experiments on three bearing datasets verify the effectiveness and superiority of the proposed method.