Efficient 2D DCNN approach for detecting and classifying faults in modular power converters

Abstract

Automatic fault detection plays an important role in increasing the dependability of modular multilevel converters (MMCs). Most existing data-driven methods rely on complex signal processing (e.g., signal segmentation) of sensor data collected from MMCs, which degrades the accuracy of the classification model due to the ignorance of the time dependencies among the various signals. Besides, such approaches require hand-crafting efficient ensemble techniques to achieve acceptable fault detection results. Differently, this paper proposes a novel 2D deep convolutional neural network (DCNN) approach for detecting and classifying faults in modular power converters. Specifically, the proposed approach employs a pre-processing stage that converts the measured signals into 2D signals (like a 2D image) and then uses a developed 2DCNN to automatically detect and classify faults in MMCs. This key advantage of the proposed method enables automatic learning of fault patterns from sensor data and precise modeling for the temporal dependencies among the measured signals. The DCNN is trained on a dataset containing diverse MMC operating conditions. Simulation results on the dataset and real-time evaluation in MATLAB Simulink demonstrate the efficiency of the proposed approach. It achieves a fault detection accuracy of 100% and a classification accuracy higher than 85%. The findings emphasize the potential of the proposed approach as a valuable tool for enhancing the reliability and performance of MMCs.