Condition Monitoring of DC-Link Capacitors Using Hidden Markov Model Supported-Convolutional Neural Network

Abstract

Non-invasive condition monitoring techniques have been developed for various electrical components within different power electronic topologies in order to increase reliability and decrease maintenance costs for these systems. DC-link capacitors are a component of particular attention for these condition monitoring systems due to their outsized effect on cost, size, and failure rate for power electronic converters. A non-invasive, online condition monitoring system is proposed in this paper which estimates the health of the MPPF DC-link capacitor within a 3-phase inverter. Current measurements are collected using a current transducer (CT) on the DC-bus, and a novel condition monitoring method of time-frequency image classification is used to analyze high frequency electromagnetic interference (EMI) content around 15-43 MHz. The proposed system uses a continuous wavelet transform (CWT), convolutional neural network (CNN), and Hidden Markov Model (HMM) to classify DC-link capacitor health into one of five stages with 99.9% accuracy.