Multiple IGBT Open-Circuit Fault Diagnosis Strategy for MMC Using Feature Reconstruction-Based Recurrent Neural Network

Abstract

The modular multilevel converter (MMC) open-circuit fault diagnosis under full power condition has become a challenging problem. Especially, when MMC operates under low-power conditions, there are significant fluctuations in capacitor voltage and the existing fault diagnosis strategies face difficulties in selecting appropriate thresholds. This article proposes a multiple open-circuit fault diagnosis strategy for MMC using the feature reconstruction-recurrent neural network (FR-RNN). In this method, the dominant characteristics of the theoretical and actual capacitor current of the submodules (SMs) are presented as the preferred feature set. The auxiliary mathematical feature based on capacitor current was screened using maximal information coefficient (MIC), and then, the selected auxiliary mathematical feature and the preferred feature set were inputted into the RNN for FR to realize fault diagnosis based on the optimal feature set. The proposed strategy can directly achieve accurate fault localization and fault-type judgment of SM faults without additional detection steps and diagnostic delay. In addition, the determined optimal feature set greatly simplifies the network computation, thus avoiding deviation caused by multidimensional and redundant features on the diagnosis outputs. Simulation and experimental studies confirm the effectiveness of strategy.