Torque Estimation and Control of PMSM Based on Deep Learning

Abstract

Aiming at the torque control precision of motor is reduced due to the variation of motor parameters and the disturbance of permanent magnets. A torque observer based on BP neural network was proposed. Firstly, the d-q axis current of the motor, the angular velocity of the motor and the rotor position are used as input layers, and the electromagnetic torque of the motor is used as an output layer to construct a torque end-to-end mapping relationship based on BP neural network. Then, the trained network topology is integrated into the Matlab/Simulink module, which is the torque observer. Finally, the torque observer is substituted into the traditional field-oriented control (FOC) strategy to verify the effectiveness and feasibility by the Simulink simulation. The simulation results show that the torque observer constructed in this paper have the advantages of independent traditional mathematical model, without influence of motor parameters and spatial harmonics, and also have good steady-state performance and transient response.