An Efficient Capsule Network Model Performance Analysis from Fast Fourier Transform Features for Prediction of Stator Winding Temperature in Permanent Magnet Synchronous Motor

An electric motor's temperature estimation is one of the critical factors governing operational efficiency, overheating prevention, and maximizing motor life. The utility of such accurate predictions will eventually lead to a priori detection of a fault, thereby improving system reliability. This study recommends a deep learning-based approach to PMSM temperature prediction. Missing data preprocessing is done by the KNN imputer and autoencoder. FFT is applied to extract the significant features in the frequency domain. Finally, these extracted features are used for training and testing of both CNN-LSTM and Capsule Network models. Results indicate that the Capsule Network surpasses the CNN-LSTM by achieving a reduced MSE of 0.0239 and a better R² score of 0.8823 in their performance on PMSM temperature prediction.