Fully Simulated Data-Driven Domain Generalized Method for Multiphase Converters Fault Diagnosis

Abstract

This article investigates the generalization capabilities of deep learning models for diagnosing faults in multiphase converter power switch devices. Traditional fault diagnosis approaches depend heavily on real-world fault data for model training. However, in industrial settings, the infrequent failures of multiphase converters and the prohibitive costs of fault experiments result in a significant scarcity of actual fault data. This limitation diminishes the reliability of models trained solely on simulation data when applied to real-world situations. To overcome this challenge, this article proposes an innovative method to improve cross-domain fault diagnosis efficacy without relying on experimental domain samples. Initially, the research employs a normalization preprocessing strategy that utilizes phase current reconstruction to minimize temporal disparities among samples. A convolutional autoencoder is then used to extract deep features from the multiphase current signals. Additionally, this article integrates deep metric learning with classification techniques to enhance the model's discrimination and clustering abilities. The key advantage of this method is that in scenarios where experimental domain data is scarce, the generalization diagnosis of open-circuit faults in multiphase motor drive systems with various parameters and types can be achieved using only simulation domain data. Furthermore, robustness and rapid fault diagnosis are realized. Experimental results and comparative analyses prove the effectiveness of the developed diagnostic algorithm. Moreover, we have made all the datasets used in this article publicly available.

Topics & Concepts