Data-Driven Robust Control for Railway Driven Independently Rotating Wheelsets Using Deep Deterministic Policy Gradient

This paper presents a data-driven robust controller for the active steering of driven independently rotating wheels (DIRW). Associated with a two-axle DIRW vehicle, a reinforcement learning controller called Deep Deterministic Policy Gradient (DDPG) is applied to improve the guidance and curve-negotiation behaviour of the DIRW system. We implement deep neural networks in DDPG to learn complex vehicle behaviours by training with data generated dynamically from non-linear simulation models. The controller can achieve adaptive optimization through online training episodes. The DDPG controller’s effectiveness is verified by the co-simulation method: the DIRW railway vehicle’s dynamics model is established in SIMPACK, and the data-driven controller is trained and deployed in MATLAB. The simulation results show that for the DIRW system, the proposed control approach can improve the IRW’s running performance and can significantly reduce the wheel-rail wear in both straight and curved tracks.