Machine learning-based switch fault localization in multilevel inverter

Multilevel inverters (MLIs) are commonly used in modern AC power applications because they generate superior voltage waveforms and lower voltage stress on the switching devices. However, the switches in these converters often face varying electrical and thermal conditions, which make them vulnerable to open-circuit faults (OCFs). Such faults can distort the output waveform and affect the overall reliability of the system. In this article, a novel fault localization method is developed for a reduced-device-count MLI to detect single and multiple switch OCFs. The proposed method requires only the output voltage of the inverter and a set of informative features extracted from it. These features include simple statistical measures along with time-frequency-based indicators obtained from the discrete wavelet transform, namely wavelet-band energy and Shannon entropy. Using these combined features, different machine learning classifiers are trained and compared. Among the tested models, the random forest classifier shows the best performance, achieving an accuracy of 99.50% for a 70:30 train-test division. The proposed scheme can locate faulted switches within about 9–25ms. The complete diagnostic setup is built and evaluated in MATLAB/Simulink and further verified using hardware-in-loop tests on the dSPACE platform, confirming its suitability for power-electronic applications.