Adaptive High-Frequency Injection Control for Low-Voltage Encoderless PMSM Drives Using Reinforcement Learning

Abstract

It is especially difficult to control low-voltage permanent magnet synchronous motors (PMSMs) in the lowest and nearest speed range because the magnitude of the back-electromotive force (back-EMF) is too low to support the traditional observer-based position sensing methods. The drawback with this approach is an inevitable trade-off between position information extraction and torque ripple, caused by using constant injection amplitudes; there is a trade-off between estimation error and torque ripple, and more harmful trade-offs involve added harmonic distortion, acoustic noise, and decreased efficiency. ) In this paper, we suggest an adaptive injection strategy of the high-frequency PMSM drive with 48 V using reinforcement learning (RL), whereby the injection amplitude is dynamically restructured depending on real-time operating conditions. Amplitude selection problem is presented in the form of a Markov decision process and a Deep Q-Network agent is trained to minimise a multi-objective cost criterion, which includes rotor position estimation error, torque ripple, and total harmonic distortion. The proposed controller is embedded into a field-oriented control system and confirmed to a full-fledged MATLAB/Simulink platform at low speed and load disturbance conditions. Compared with traditional fixed-amplitude and heuristic adaptive control, the comparative results with improved amelioration of faster convergence, torque ripple, and harmonic distortion are seen with minimal harmonic distortion and excellent rotor position observability. The results prove that learning-based amplitude adaptation is an effective method to overcome the accuracyversesripple trade-off in HFI-based sensorless PMSM motors, so the technique is suitable in locations of small-scale 48-V traction and automotive auxiliary systems that need high performance at lower speed.