Transferable Neural Network Method Used for PMSM Permanent Magnet Temperature Estimation in Electrical Drive System

Nowadays, electric drive systems are increasingly used in many industrial scenarios, especially in electric vehicles (EVs). Many permanent magnet synchronous motors (PMSMs) lack stable and accurate temperature monitoring and can only rely on expensive motor designs to ensure safe operation at high temperatures. Temperature monitoring using traditional thermal modeling methods requires workers to have strong expertise, such as the selection of a priori parameters. This paper proposes a transferable estimation method for the temperature of the permanent magnets (PMs) of the motor rotor. The domain adaptation module is added to the 1-D convolutional neural network (CNN) to realize the transfer function of the training model. Before the model pre-training stage, the generative adversarial network (GAN) technique is applied to data augmentation for improving the generalization ability of the pre-trained model. Since the training of the domain adaptation module does not require labeled data, the model can be migrated across multiple different motors and more accurately estimate the temperature of the PMs. The final experimental part verifies the accuracy of this method through the professional measured temperature data of the PMSM. The mean square errors of proposed method are under 0.51.