Interests and Limits of Machine Learning-Based Neural Networks for Rotor Position Estimation in EV Traction Drives

Abstract

In this paper, a novel rotor position estimator for an interior permanent magnet synchronous motor is presented and evaluated. The proposed estimator lies on one of the most popular methods in the field of artificial intelligence: Machine learning-based neural networks algorithm. The main interest is to introduce a cost-efficient position estimator that is comparable to classic methods in terms of functional performances. The estimator model is built by learning from a dataset that associates phase currents and voltages to the rotor position. Learning signals are generated using a simulation model. This is primarily intended to save the resources invested in testbench trials. In this work, offline training steps and results are described and commented. The efficiency of the proposed position estimator is first verified by functional simulations. Second, real-time experiments are conducted on an actual scale testbench. The NN-based estimator covers a wide speed range and is implemented in the context of IPMSM-based EV traction drives. More broadly, these findings can also be applicable to the ac-based electric drives for the position estimation purpose.