Learning Thermal Properties and Temperature Models of Electric Motors with Neural Ordinary Differential Equations

The importance of thermal monitoring in electric traction drives gains more and more momentum as overload capabilities have to be utilized increasingly to stay competitive in the electric vehicle industry. Among real-time capable algorithms for the temperature estimation of in-series-unaccessible motor components, lumped-parameter thermal networks denote a popular choice but require extensive expert knowledge and motor data sheet information for their design. In this work, it is shown that with emerging techniques from the cross section of ma-chine learning and physical modeling of heat transfer with differential equations, new chances for accurate real-time capable thermal modeling on a completely measurement-data-driven base arise. A data-efficient training scheme for thermal neural networks and their generalization, neural ordinary differential equations, is presented that converges to cross-validated estimation accuracies of less than 2K2 mean squared error, and below 6 K worst-case error at only 149 model parameters.