Switch fault identification scheme based on machine learning algorithms for PV-Fed three-phase neutral point clamped inverter

Abstract

Any faults in power electronic switches must be detected and located promptly to certify the unchanging operation and reliability of multilevel inverter (MLI) systems. Machine learning (ML) techniques are increasingly being utilized to diagnose faults in MLIs due to their advantages, such as high precision, less calculation time, and reduced complexity. A faulty switch identification and phase identification method based on different ML classifiers is presented in this paper, explicitly targeting open-circuit faults in switches in a PV-fed 3-phase three-level Neutral Point Clamped (NPC) inverter. The NPC MLI output voltage and current signals are input signals to classify faults. The information or essential feature from the raw voltage and current data is extracted using a recursive discrete Fourier (RDF) followed by a discrete wavelet transform (DWT). Finally, the standard deviations of the approximate and detailed coefficients are used as input for ML algorithms. To do the comparative analysis, a variety of ML techniques are then applied to these features, including logistic regression (LR), random forest (RF), quadratic discriminate analysis (QDA), K-nearest neighbour (K-NN), and support vector machines (SVM). The experiment results indicate that the RF algorithms achieve the highest classification accuracy of 99.59 % for various switch fault conditions.