A multi-source learning method for open-switch fault diagnosis in power converters

Abstract

This dissertation explores a novel multi-source domain adaptation extreme learning machine (MDAELM) method for diagnosing open-circuit faults in insulated gate bipolar transistors (IGBTs) within three-phase inverters. Traditional fault diagnosis methods often fail to address challenges arising from data distribution shifts and domain variability in real-world applications. To overcome these limitations, this research employs maximum mean discrepancy (MMD) to align data distributions across domains and introduces a soft-label weighted voting mechanism to enhance classification accuracy. Experimental results demonstrate that MDAELM outperforms conventional methods, such as single-domain extreme learning machines (ELM) and support vector machines (SVM), in terms of fault classification accuracy, robustness, and computational efficiency. This study provides a scalable and effective solution for power converter fault diagnosis, offering potential applications in industrial systems with varying operational conditions.