Reinforcement Learning-Based Vibration Control for Half-Car Active Suspension Considering Unknown Dynamics and Preset Convergence Rate

Abstract

Electromagnetic actuators, characterized by their lack of pneumatic or hydraulic circuits, rapid response, and ease of control, have the potential to significantly enhance the dynamic performance of automotive active suspensions. However, the complexity associated with their models and the calibration of control parameters hampers the efficiency of control design. To address this issue, this paper proposes a reinforcement learning vibration control strategy for electromagnetic active suspension. Firstly, a half-car dynamic model with electromagnetic active suspension is established. Considering the unknown dynamics of the actuator and its preset convergence performance, an optimal control method based on reinforcement learning is investigated. Secondly, a heuristic PI adaptive dynamic programming algorithm is presented. This method can update to the optimal control solution without requiring model parameters or initial design parameters. Finally, the energy consumption and dynamic performance of this method are analyzed through rapid prototyping control simulation. The results show that the ride comfort of the vehicle suspension can be improved with the given preset convergence rate.