Physics-informed reinforcement learning for adaptive high-frequency injection in encoderless low-voltage PMSM drives

Abstract

It is difficult to control permanent magnet synchronous motor (PMSM) drives running at extra-low voltages with encoderless control because the back-EMF signal to estimate rotor position is weak, and this requires the injection of high-frequency (HF) signals. Traditional methods use constant or manually tuned injection levels, and these tend to cause large torque ripple, inaccurate estimation when under dynamic loading, and an inability to counteract parameter drift. The paper is related to the issue of online optimal HF injection amplitude choice in the encoderless 48 V PMSM drives and proposes a physics-inspired reinforcement learning (PIRL) system. This is aimed at obtaining the right low-speed positioning and reducing the torque ripple and power losses on different operating conditions. The suggested approach incorporates directly into the reinforcement learning reward terms the PMSM electromagnetic voltage equations, which restrict exploration to physically consistent space and enhance stability in the learning process. The PIRL agent is trained in a deep deterministic policy gradient architecture in a MATLAB/Simulink-Python co-simulation environment, based on which the PIRL agent adjusts the injection amplitude of HF in real time. Simulation outcomes show that the suggested methodology reaches approximately four times faster convergence with conventional reinforcement learning and reaches up to 65 percent of torque ripple reduction without a disturbed position estimation when operated in a speed range of 0-500 rpm. The findings show that physics-informed learning offers an efficient and energy-saving solution to adaptive encoderless control in extra-low-voltage PMSM drives, which has better resilience to changes in parameters with a low computational cost.