Physics-Constrained Neural Network-Based Load Torque Observer for Traction Electric Machine

A torque observer is essential to prevent torque derating under thermal stress. It accurately monitors the traction Electric Machine (EM) operating temperature and adjusts torque estimation accordingly. By accounting for temperature variations, it ensures that the EM operates within its safe thermal limits, preventing overheating and maintaining consistent performance. This paper compares the design of a load torque observer based on a physics-constrained neural network (PCNN) with a classical neural network (NN) based torque observer. The PCNN-based torque observer accurately estimates EM torque while accounting for temperature effects but NN based torque observer fails. Furthermore, stability analysis of PCNN torque observer for electric traction machine control is established. The proposed observer is validated and results show its accuracy of about 98% which remains the same in the validation process whereas the accuracy of the classical neural network is affected by parameter variation.