Deep Residual Convolutional and Recurrent Neural Networks for Temperature Estimation in Permanent Magnet Synchronous Motors

Most traction drive applications using permanent magnet synchronous motors (PMSMs) lack accurate temperature monitoring capabilities so that safe operation is ensured through expensive, oversized materials at the cost of its effective utilization. Classic thermal modeling is conducted with e.g. lumped-parameter thermal networks (LPTNs), which help to estimate internal component temperatures rather precisely but also require expertise in choosing model parameters and lack physical interpretability as soon as their degrees of freedom are curtailed in order to meet the real-time requirement. In this work, deep recurrent and convolutional neural networks with residual connections are empirically evaluated for their feasibility on the sequence learning task of predicting latent high-dynamic temperatures inside PMSMs, which, to the authors' best knowledge, has not been elaborated in previous literature. In a highly utilized PMSM for electric vehicle applications, the temperature profile in the stator teeth, winding, and yoke as well as the rotor's permanent magnets are modeled while their ground truth is available as test bench data. A model hyperparameter search is conducted sequentially via Bayesian optimization across different random number generator seeds in order to evaluate model training consistency and to find promising topologies as well as optimization strategies systematically. It has been found that the mean squared error and maximum absolute deviation performances of both, deep recurrent and convolutional neural networks with residual connections, meet those of LPTNs, without requiring domain expertise for their design. Code is available at [1] to assist related work.