Machine-Learning Approach to Model Junction Temperatures in Automotive Inverters

Abstract:
Increasing power density of automotive inverters lead to an increasing demand for accurate lifetime and reliability models. As such models are closely dependent on junction temperatures, they benefit from accurate temperature estimation methods. In this contribution, a machine-learning approach to model semiconductor junction temperatures is presented. A linear regression model (Ordinary Least Squares) and a simple neural network (Multilayer Perceptron) were investigated. The models were trained and evaluated with data from a test bench incorporating a 1200 V SiC power module. It is demonstrated, that even linear models show a predictive performance of below 1 K² mean squared error on the designed test set, as long as the input data is preprocessed with exponentially weighted moving averages (EWMAs) and non-linear features are supplied. The data generation, feature engineering, hyperparameter optimization, model training and performance as well as benefits and limitations are shown and discussed.