Improved Nonlinear Estimation in Thermal Networks Using Machine Learning

Emerging new technologies as found in modern electric cars must compete with existing technology in terms of quality and price. The pressure on the price is especially high in the automotive section. Research in the field of state estimation is of high potential for reducing the number of sensors, thus enabling cost savings in production. The methods of machine learning are also increasingly influencing this field of research. This article focuses on the thermal behavior of fluid cooled automotive IGBT (insulated gate bi-polar transistor) inverters and the application of machine learning methods in estimation tasks in nonlinear thermal networks. For this purpose, a parameterized grey-box model is designed using a linear thermal Cauer network in combination with numerical parameter fitting. Special emphasis is put on regression methods that are used to fit nonlinear thermal resistances to measurement data. An unscented Kalman filter (UKF) is applied to estimate states of the thermal network. In addition, a feed-forward artificial neural network (ANN) is trained on the estimation error using sensor signals as predictors to improve the estimation. Results on measurement data from a test bench show a significant improvement by the methods.