This paper focuses on developing a torque and stator temperature prediction model for permanent magnet synchronous motors using neural networks. The model can predict torque and four other temperature parameters at the permanent magnet surface, stator's yoke, tooth, and winding. The motor's torque and temperatures are predicted without installing any additional sensors into it. Using the training dataset with Levenberg-Marquardt optimization and Bayesian regularization algorithms, the predicted model has the best performance with the least mean square error and the best $R^{2}$ values. Also, the prediction of testing data shows that the estimated model follows closely with actual values. This is true for all the five output parameters.

Torque and Temperature Prediction for Permanent Magnet Synchronous Motor Using Neural Networks

The ubiquitous adoption of permanent magnet synchronous motors (PMSMs) as the electric motors of choice for traction drive applications especially in manufacturing and the electric vehicle industries birthed the need for monitoring the temperatures of its critical components to control the effects of overheating. In proffering solutions, several techniques have been employed by researchers spanning decades. These include the sensor-based method, methods based on classic thermal theory, electric circuit theory and the hybrid lumped-parameter thermal networks (LPTNs). These however have deficiencies ranging from requiring expertise for efficient modelling to one or the other of lacking interpretability and not meeting reliability requirements. Recent studies have seen an increased application of machine learning techniques to other fields like healthcare with convincing results. In this work, several machine learning (ML) models were evaluated on their estimation error after training on test bench data from a PMSM for the task of predicting the temperatures of the rotor, stator yoke, stator tooth and stator winding. Diverse regression algorithms were applied and include linear regression (LR), k-nearest neighbours (kNN) regression, random forest (RF) and decision tree (DT). It is observed that the stator yoke records the least error of prediction while the pm records the highest and in general, the stator components record the least error compared to the rotor component.
Keywords- permanent magnet synchronous motors, machine learning, linear regression, temperature estimation, random forests

Kenneth Anuforo

Papers

Temperature Estimation in Permanent Magnet Synchronous Motor (PMSM) Components using Machine Learning