A New Weighted Mechanism-Based Partial Transfer Fault Diagnosis Method for Voltage Source Inverter

Abstract

Transfer learning is a promising technique used to deal with the issue of domain transfer in fault diagnosis of three-phase PWM-VSI. However, there is limited research on partial transfer for VSI, where the data category of the source domain is the complete health conditions and contains the data category of the target domain. Consequently, this article designs a new weighted mechanism to solve the open-circuit fault diagnosis for VSI in three-phase permanent-magnet synchronous motor (PMSM) under partial transfer scenarios. The distribution difference between the two domains is decreased by minimizing the Wasserstein distance between the weighted source domain data and target domain data, while the weights are obtained in the process. The outlier data in the source domain are assigned with small weights to diminish their effect on the classification results, thereby enhancing the performance of the model for partial transfer diagnostic. Simultaneously, the transferable recognition network is trained by the weighted cross-entropy loss on the source domain data and the conditional entropy loss on the target domain data. A training strategy is proposed to optimize weights and network parameters alternately, where one aspect is fixed while the other is updated iteratively. The proposed method is verified by experiments.