Adaptive endocrine PID control for active suspension based on reinforcement learning

Abstract

Recent research has focused on active suspension systems because of their real-time ability to adapt to a variety of road surfaces, external perturbations, and potential to control the smoothness of vehicles. Since the parameters of conventional PID controllers used for active suspension control have limitations due to their inability to adapt to external changes, the fuzzy PID controllers are developed to overcome such limitations. However, fuzzy control have certain disadvantages related to the manual definition of fuzzy rules and variables. In this paper, a first step in improving the robustness of the PID control is to use the endocrine framework, which is considered to be highly nonlinear and complex in the vehicle suspension system. A deep reinforcement learning algorithm is then used to train the intelligence to provide an efficient strategy for adaptive gain adjustment for the endocrine PID, which requires no prior knowledge of active suspension control. The dynamics of the whole vehicle are modeled using ADAMS to analyze the dynamic characteristics of the vehicle at different speeds and road surfaces. The results show that the active suspension based on deep reinforcement learning controlling reduces vertical acceleration of the body more effectively and improves ride comfort more efficiently without sacrificing dynamic suspension deflection or dynamic tire load as compared with passive suspension or fuzzy PID suspensions. Further, the controller performs well under conditions such as changing road grades and vehicle speeds, indicating a good generalization.