Multibody dynamic modelling and guided reinforcement learning for a novel air suspension system of semi-trailers

Abstract

Vehicle vibrations during long-haul logistics transportation can significantly degrade ride comfort and increase the risk of cargo damage. To address this issue, this paper proposes a novel torsion-beam air suspension configuration and a guided reinforcement learning–based semi-active control strategy. First, a multibody dynamics model of the proposed torsion-beam air suspension is established in ADAMS and validated against full-scale road tests, confirming its ability to accurately reproduce the dynamic behaviour of the suspension under real operating conditions. To improve the adaptability of semi-active suspension control under complex and varying road excitations, a guided reinforcement learning control framework is subsequently developed by integrating model predictive control (MPC) with the deep deterministic policy gradient (DDPG) algorithm. In this framework, the predictive optimisation results generated by MPC are incorporated into the reinforcement learning process through reward shaping, constraining the policy search within physically feasible regions and enabling coordinated optimisation of the semi-active damper. The effectiveness of the proposed suspension configuration and control strategy is evaluated through ADAMS–Simulink co-simulation and hardware-in-the-loop (HiL) experiments. HiL test results on a C-class road demonstrate that the proposed G-DDPG controller reduces the RMS values of sprung-mass acceleration, pitch angle, roll angle, and tire dynamic load by 14.11%, 16.15%, 20.30%, and 19.39%, respectively, relative to the passive suspension, demonstrating superior vibration attenuation performance and improved attitude stability compared with both MPC and standard DDPG controllers.