A New Machine Learning Based Approach for Aluminium Electrolytic Capacitors Health Status Monitoring

Power converters are one of the main elements of numerous critical systems. These converters are composed of several components, some of which are particularly vulnerable, namely semiconductors and capacitors. Therefore, monitoring those components' condition is fundamental, mainly in applications that require high reliability and safety. One of the most commonly used capacitor in power converters is the aluminum electrolytic capacitor (AEC), whose aging manifests itself through an increase in its equivalent series resistance (ESR) and a reduction in its capacitance (C). Several techniques have been proposed to assess the AEC state of health, having the Equivalent Series Resistance (ESR) value as a reference. For this purpose, a current sensor is used in the capacitor branch. However, the current sensor imposes practical restrictions; namely, it requires space and introduces an unwanted resistive and inductive effect. In this paper, a Decision Tree Regressor (DTR) based model, combined with the Short Time Least Squares Prony's (STLSP) approach is used to detect the fault. At first, the STLSP algorithm is used to train the DTR model to predict the capacitor current amplitude (A-Ic). Then, using both STLSP and DTR algorithms, it is possible to estimate the ESR value without the need to use a current sensor inside the converter.