Machine learning-based prediction of on-state voltage for real-time health monitoring of IGBT

Abstract

Real-time temperature monitoring of power semiconductors is a powerful tool for the predictive maintenance and lifetime prediction of power converters, and is typically performed by measuring temperature sensitive electrical parameters (TSEP). In this work, we demonstrate a machine learning-based method to estimate the on-state voltage of a real converter prototype featuring variable load of electric vehicle (EV) and photovoltaic (PV) systems under different ambient temperatures induced in a climate chamber. The method provides for the first-time insights on the uncertainties and feature importance for the predictions, and is aimed to be industrial compatible by applying only methods that are well established in industry. The approach is further flexible to any converter system, independently of its specifications. We have used the Support Vector Machine, K-Nearest Neighbors, and Decision Tree algorithms to predict the on-state voltage as functions of the readily measured parameters of a converter. We show that for the PV case, the K-Nearest Neighbor method yields the lowest mean absolute error of about 0.75 % for prediction, while for EV, the K-Nearest Neighbor algorithm gives the lowest mean absolute error of 5 %.

Keywords

• Condition monitoring
• Machine learning
• Power electronic converters
• Predictive maintenance
• Si IGBT