Deep reinforcement learning-based smart vibration control for magnetorheological suspension considering nonlinear dynamics

Abstract In response to the complex nonlinear characteristics of magnetorheological (MR) dampers, which make it difficult to achieve strict vibration reduction requirements in time-varying working environments due to uncertainties, this paper aims to develop a refined model for the MR damper and improve the deep reinforcement learning network based on its dynamic characteristics to optimize the control strategy. Firstly, a refined dynamic model for the MR damper is designed, and its Simulink model is established and corrected based on the characteristic data of the custom damper. Secondly, an optimization framework for suspension control is constructed, where the suspension and road surface serve as the environment, and the proximal policy optimization (PPO) algorithm is used as the agent. Next, the nonlinear relationship between the damping force of the damper and its performance indicators is explored, and a PPO-neural ordinary differential equations (NODEs) algorithm is proposed by combining NODE with PPO. This approach allows for continuous fitting of state variables, enabling high-precision nonlinear fitting between the state space and policy representation. Finally, a quarter-car test bench is set up using the custom MR damper, and the simulation and experimental results of passive suspension, fuzzy PID, original PPO, and the proposed method are compared. The experimental results show that the proposed method demonstrates significant advantages: under class C road conditions, the body acceleration in simulation and test bench experiments is reduced by 30.22% and 77.51% respectively compared to passive suspension; 20.55% and 37.82% compared to fuzzy PID; and 11.26% and 25.53% compared to the original PPO. Additionally, the simulation results on Class B random road further validate the good generalization performance of the proposed method.