Vibration Suppression of Active Magnetic Bearings Based on Reinforcement Learning Optimized ADRC

Abstract

To enhance the robustness and stability of active magnetic bearing (AMB) rotor systems under complex disturbances and nonlinear operating conditions, this paper proposes an adaptive active disturbance rejection control (RLADRC) strategy based on reinforcement learning. Traditional ADRC compensates for unbalanced forces through an extended state observer (ESO), but its key parameters (observer gains) rely on empirical tuning, making it difficult to adapt to dynamically changing environments. This study introduces reinforcement learning algorithms to optimize observer gains online through real-time interaction with the system environment, thereby achieving adaptive optimal estimation and compensation for unbalanced disturbances. Simulation results indicate that, compared with traditional ADRC, the proposed method reduces the amplitude of unbalanced forces in each degree of freedom by 90% at a constant speed and results in smaller rotor vibrations when subjected to external impacts. Experimental verification demonstrates that the controller significantly improves anti-interference capability compared to traditional PID control.