Application of Thermal Neural Networks on a Small-Scale Electric Motor

Abstract:
In the automotive market and automation industry, the race for ever increasing power and energy densities of electric motors leads to an endeavor of developing real-time-capable thermal models. The precise thermal monitoring of critical motor components promises material reduction and higher material utilization. Among classical, expert-based, lumped-parameter thermal networks (LPTNs) and black-box, data-driven machine learning approaches, synergies have been identified in the recent literature. In this work, one of these synergies – thermal neural networks (TNN) – are evaluated on a test bench equipped with a prototypical 110Wpermanent magnet synchronous motor that features thermal sensors. The demonstrated cross-validation score of the TNN with an average mean squared error of 0.48 K² and absolute estimation errors of under 2 ◦C for 98.6% of all samples excels a data-driven classic LPTN that acts as baseline. The TNN features roughly five times more parameters than the expert-based LPTN, but is optimized in a fraction of the time with no geometry or material information involved.