Induction Motor Temperature Prediction via Transfer Learning on Physics-Based Synthetic Data Using Deep Neural Networks

This work addresses the challenge of accurately predicting steady-state temperatures in induction motors, particularly when limited experimental data is available. A hybrid methodology is proposed, combining physics-based modeling with deep learning. An LPTN model, derived from FEM analysis of a 3hp motor, is used to generate synthetic data representing thermal behavior under varying loss conditions. A deep neural network is trained on this dataset and adapted to a 15hp motor using transfer learning with only three experimental samples. A calibration function is further introduced to correct rotor temperature prediction errors.