Fault Diagnostic Techniques using Deep Learning on Power Electronics Systems Integrations

Power electronic converters are fundamental components in modern industrial drive systems, renewable energy interfaces, electric transportation, and energy storage applications. As converter topologies become increasingly complex, reliable and efficient fault diagnosis methods are required to ensure safe operation and reduce maintenance costs. Conventional model-based diagnostic approaches often face limitations under parameter uncertainty and varying operating conditions. Data-driven and artificial intelligence (AI)-based techniques provide a flexible alternative by learning fault characteristics directly from measured signals. This thesis investigates AI-based fault diagnosis and condition monitoring strategies for power electronic converters, with emphasis on scalability, robustness, and practical implementation. Initial studies evaluate shallow neural networks and classical machine learning methods for open-circuit fault detection in voltage source inverters. The main contribution is the development of hierarchical convolutional neural network (CNN) and attention-based architectures for fault localization in Cascaded H-Bridge converters, enabling structured identification of faults at phase, module, and switch levels. Comparative analysis shows that hierarchical CNN models offer compact solutions, while attention-based models provide improved scalability for higher-resolution localization tasks. The methodology is further extended to Dual Active Bridge converters, demonstrating the adaptability of AI-based diagnostics to isolated DC--DC topologies. In addition, AI-based thermal modeling approaches are investigated for condition monitoring applications. Overall, the results demonstrate that compact and structured AI models can provide accurate and computationally efficient diagnostic solutions for modern power electronic systems.