Rahman Dashti

Background: Modular Multilevel Converters (MMCs) are widely used in highvoltage drives and power transmission systems due to their modular architecture, high efficiency, and excellent output power quality. However, open-circuit faults in submodules can cause system instability, current distortion, and reduced reliability. Consequently, developing accurate, fast, and generalizable fault-detection methods is essential. Aim: This research aims to develop a two-stage deep-learning-based framework for detecting open-circuit faults in MMCs, achieving high accuracy in identifying both the occurrence and precise location of faults across all submodules. Methodology: All fault scenarios were comprehensively simulated in MATLAB/Simulink, and the resulting data matrices were exported to the coding environment. In both stages, measured signals were transformed into time– frequency representations using the Synchrosqueezing Wavelet Transform (SSWT). These representations were then denoised, processed with adaptive thresholding, and converted into images to serve as inputs to Deep Convolutional Neural Networks (DCNNs). In the first stage, circulating currents and output currents were used to classify seven scenarios (one fault-free and six faulty), represented as 150×150 pixel images. The network achieved an accuracy of 99.29%. The second-stage network was trained on 30 scenarios for the positive A half arm submodules, achieving 99.30% accuracy. This system can be used for Negative A submodules as well but for two other phases DCNN needs to be trained for 30 different open-circuit scenarios again for their positive half arm and similarly uses for negative half arms