Deep Neural Network Modeling for Accurate Electric Motor Temperature Prediction

Abstract

Electric motors are becoming widely used in many different applications, such as electric cars, and turbines. Measuring the temperature of internal components of an electric motor, like permanent magnet synchronous motor (PMSM) is vital to maintain its safe operation. However, measuring the temperature of the permanent magnet and stator directly comes at the expense of higher cost and additional hardware requirement, for instance, a sensor network. To overcome these limitations, machine learning (ML) techniques can be employed to model the mentioned parameters without the need of specialized sensors and design ideas for housing them inside the motors. Classical methods, like lumped-parameter thermal networks (LPTNs) are capable of calculating the temperature of internal elements of PMSMs. But, these methods require expertise and may lack an acceptable accuracy. In this study, two deep neural networks (DNNs) were modeled using convolutional neural network (CNN) and long short-term memory (LSTM) units to predict the temperature of four target values of PMSMs: stator tooth, stator yoke, stator winding, and permanent magnet. For attribute conditioning, exponentially weighted moving average (EWMA) and exponentially weighted moving standard deviation (EWMS) were applied. A thorough ablation analysis shows that the CNN-based model predicts the targets better than the LSTM model with an average mean squared error (MSE) of 2.64 ${\circ} \mathrm{C}^{2}$ and an average R <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> of 0.9924. It is also found that the proposed CNN-based model achieves a 13% mean average performance (mAP) improvement compared to the existing state-of-the-art solution.