PMSM Traction-Motor Thermal Digital Twin Using LPTN and Machine Learning

An electric-vehicle traction motor must not overheat, but the rotor magnets cannot be measured in a production car. This project builds a virtual temperature sensor, a thermal digital twin, from current, voltage, speed, torque and coolant temperature, and compares identified lumped-parameter thermal networks (LPTN) with EWMA + ridge, EWMA + MLP, a dilated 1-D CNN and a thermal neural network on measurement profiles held out whole (60/62/74) from the Paderborn data (1,330,816 rows, 69 profiles). Lowest MSE per node among all ten estimators: magnet: Thermal neural network (2.61 K2); winding: Thermal neural network (2.61 K2); tooth: EWMA + MLP (2.08 K2); yoke: Thermal neural network (1.18 K2). Among the models that need no start temperature the lowest MSE is: magnet EWMA + MLP (4.64); winding EWMA + MLP (5.09); tooth EWMA + MLP (2.08); yoke EWMA + MLP (1.92): EWMA + MLP and ridge lead. A closed-loop torque- derating scenario is a simulation, not a vehicle test.

Category: Automobile Engineering

Automobile Engineering Project AU04 from ProTechSC Academy, with graphical abstract, video demonstration and project guidance.