Abstract Summary
The magnetic suspension control system is the key component of the maglev train to ensure the suspension gap between the maglev train and its guideway at a stable value. In general, the maglev car body is supported by several bogies which consist of two suspension modules. The two suspension modules are decoupled mechanically by a well-designed anti-rolling beam making it the fundamental unit of maglev train. Each suspension module contains two suspension control points which are controlled separately by two individual controllers. The current single-suspension control system neglects the coupling disturbance between the two levitation points causing a conservative dynamic performance. Besides, the most common adopted linear suspension controller nowadays has met the basic requirements of engineering application, but some problems occur in the suspension system during complex working conditions and long-term passenger service which seriously affect the stability and reliability of the suspension system and even cause the partial suspension-point failure of the vehicle, which affects the safety and stability of the maglev train. In this paper, an adaptive multi-agent deep reinforcement learning (MADRL) approach for cooperative control of nonlinear maglev suspension system developed, enabling automatically adjust the control strategy through the interaction with the suspension system. The nonlinear state space of two-point suspension control system is established as an agent environment to interact with the developed MADRL control model. Multi-agent deep deterministic policy gradient (MADDPG) in MADRL is adopted to find the agents which maximize the total reward by resembling the control index for optimal control. The uncertainty factors like mass change and disturbance force in the methodological framework of the maglev system are considered in the established method. The effectiveness of the proposed method is verified by comparing with conventional PID controller through simulation.
