Open-Circuit Fault Diagnosis in Three-Phase Modular Multilevel Converters via Multi-Scale Dilated CNN with Efficient Channel Attention

Precise and timely detection of open-circuit (OC) faults in Modular Multilevel Converters (MMCs) is essential for maintaining system reliability and preventing catastrophic failures. This paper presents an optimized diagnostic framework based on a Multi-Scale Dilated Convolutional Neural Network with Efficient Channel Attention (MDCNN-ECA). Unlike traditional methods that depend on complex multi-source data, the proposed approach exclusively utilizes raw arm current signals, simplifying the measurement hardware while ensuring high precision. The architecture integrates multiscale dilated convolutions to capture temporal features across varying resolutions, coupled with an ECA mechanism that adaptively enhances sensitivity to fault-specific patterns. Despite a significant ۳-fold reduction in parameter count compared to conventional CNNs, the MDCNN-ECA achieves superior classification performance. The model was evaluated using a comprehensive dataset of ۱۳ classes, covering normal operation and ۱۲ switch-level fault scenarios. Experimental results demonstrate a diagnostic accuracy of ۹۹.۵۶%, with exceptional robustness against noise and varying load conditions. These findings confirm that the MDCNN-ECA provides a lightweight, efficient, and robust solution for real-time industrial MMC fault diagnosis.

کلیدواژه ها:

Modular multilevel converter ، Fault diagnosis ، one-dimensional convolutional neural network ، Multi-Scale Dilated Convolutional Neural Network ، Efficient Channel Attention