A Novel Model Calibration Method for Active Magnetic Bearing Based on Deep Reinforcement Learning

Abstract

Active magnetically suspended control moment gyro is a novel attitude control actuator for satellites. It is mainly composed of rotor, active magnetic bearing (AMB) and motor. As a crucial supporting component of control moment gyro, the performance of AMB is directly related to the stability of the rotor system and pointing precision of the satellites. Therefore, calibrating the parameters of AMB is essential for the realization of super-quiet satellites. This paper proposed a model calibration method, known as the deep reinforcement learning-based model calibration frame (DRLMC). First, the dynamics of magnetic bearing with damage degradation over its life cycle are modeled. Subsequently, the calibration process is formulated as a Markov Decision Process (MDP), and reinforcement learning (RL) is employed to infer the degradation parameters. In addition, experience replay and target network update mechanism are introduced to guarantee stability. Simulation results demonstrate that the proposed method identifies force-current factor of AMB during its degradation process effectively. Furthermore, additional experiments confirm the robustness of the DRLMC approach.