A Novel Compact Hierarchical Deep Convolutional Neural Network Architecture for Faulty Switch Detection and Localization in Power Converters

This paper introduces a hierarchical fault diagnostic framework based on Convolutional Neural Network (CNN) models to identify, localize, and classify Open Circuit Faults (OCFs) in Cascaded H-Bridge (CHB) power converters. The framework is organized in three stages: overall health detection, fault localization at the phase level, and switch/module fault identification. Two hierarchical topologies are investigated. In a 3-level CHB, the final stage performs switch-level localization, since each phase leg consists of a single H-bridge. In a 7- level CHB, the concept extends to module-level localization, where each module is considered as a complete MOSFET unit. Key contributions include the development and testing of multiple classifiers, namely a 2-class system health classifier, a 3-class phase-localization classifier, and switch/modulelocalization classifiers depending on the CHB topology. The proposed models achieved high prediction accuracy, with the 2-class classifier reaching 99.98% and the switch-localization classifiers in the 3-level CHB exceeding 95%. In the 7-level case, a hierarchical approach combining CNN-based phase localization with a rule-based module-localizer reached 100% module-level accuracy with an average inference time of 2.687ms and no trainable parameters. Finally, the framework was quantitatively benchmarked against conventional 1-D CNN, 2-D CNN, Fusion CNN, and classical machine learning methods (k-NN, Decision Tree, LDA, Naive Bayes). The comparative analysis highlights the trade-offs in accuracy, complexity, and real-time suitability, demonstrating the scalability and practical relevance of the proposed hierarchical approach. The methodology was validated using experimental and simulated data, including spectrograms generated by short-time Fourier transform (STFT) analysis of both healthy and faulty system signals.