Accurate Short-Horizon Multi-Target Prediction of PMSM Operational Parameters via Residual Dilated 1D Convolutional Neural Networks

Abstract

Accurate short-horizon prediction of key operating parameters in Permanent Magnet Synchronous Motors (PMSMs) is essential for ensuring operational safety, optimizing control strategies, and preventing thermal stress-induced failures. This study presents a residual dilated one-dimensional convolutional neural network (1D-CNN) framework for the simultaneous estimation of three target variables motor speed, stator yoke temperature, and stator winding temperature using a publicly available high-resolution multi-sensor PMSM dataset collected on a laboratory test bench at Paderborn University. The dataset comprises 1,330,816 samples of 13 variables without missing values and was processed through a systematic pipeline including normalization, sliding-window sequence generation (window size: 256), and train–test splitting. The proposed architecture integrates dilated convolutional layers to expand the temporal receptive field, residual connections to facilitate gradient flow, and dense layers for multi-output regression. Experimental evaluations using MSE, RMSE, MAE, and R² metrics demonstrated high prediction accuracy, achieving R² values of 0.9969, 0.9819, and 0.9698 for motor speed, stator yoke temperature, and stator winding temperature, respectively, with an average R² of 0.9829 and MAE of 26.35. Comparative feature importance analysis across three independent methods consistently identified coolant temperature, d-axis current, and ambient temperature as the most influential predictors. Residual distribution analysis confirmed low bias and symmetric error patterns across all targets. The proposed approach offers a robust and computationally efficient solution for real-time PMSM monitoring, predictive control, and condition-based maintenance.