Model Calibration of Active Magnetic Bearing Using Deep Reinforcement Learning

Active magnetic bearings (AMBs) have been widely used in rotating machinery owning to their numerous advantages including no friction and adjustable bearing dynamics. Calibration of AMB model parameters is of significant importance in control optimization, condition monitoring and fault diagnosis. Traditional calibration methods mostly based on sampling and statistical methods, which suffer from applicability and accuracy when applying to AMB. Therefore, this paper proposed an AMB model calibration method based on deep reinforcement learning. Calibration is modelled as a parameter tracking problem, the agent interacts with dynamic model to learn calibration strategies directly. Results showed that our method can accurately identify force-current factor of AMB with high robustness to model uncertainty and measurement noise. Main contributions include that model calibration was modeled as a Markov decision process (MDP), which does not require any labeled data and ground truth degradation parameters, and deep reinforcement learning was introduced to learn calibration strategy directly, guaranteeing high robustness and adaptability to different scenes.