Reinforcement learning control of lateral vibration of high-speed train carbody

Abstract Increasing speed is one of the effective ways to raise high-speed train (HST) transportation capacity. However, increasing speed leads to greater vibration of HST with passive suspension, which can reduce the smoothness and safety of HST. Thus, a twin delayed deep deterministic policy gradient (TD3) control strategy was designed to reduce the lateral vibration of the HST carbody. The 3-degree-of-freedom (DOF) and 17-DOF models were utilized in designing the TD3 controller and simulation, respectively. In addition, the ballastless track spectrum of the Chinese high-speed railway was utilized to obtain the irregularities of the track. A reward function containing cost and penalty was designed to raise TD3 controller training efficiency. A linear quadratic optimal cost function and a step penalty function were designed, respectively. The control effect of the TD3 strategy was studied by simulation. The results indicate that TD3 can effectively reduce root-mean-square (RMS) and peak values for lateral acceleration at the front and rear ends of the HST carbody, and significantly improve the lateral smoothness of the HST carbody, which verifies validity of TD3 strategy.