Vehicle Suspension Control using Physics Guided Machine Learning

Vehicle suspension systems are crucial in optimizing the vehicle's ride comfort and road holding properties. Semi-active and active suspension systems play a significant role in bridging the gap in achieving the desired vertical dynamic characteristics of the vehicles compared to the traditional non-controllable and controllable suspension systems. Conventional controllable suspension systems utilize either a completely data-driven approach toward developing a control function or a classical control framework that enables the variation of the damping characteristics of the suspension system. These approaches suffer from the volatile nature of the driving conditions due to variations in speed, tire load, road surface, road grade, banking angles, etc. In this paper, a novel approach toward the control of the vertical dynamic characteristics of a vehicle has been proposed based on a fusion of theoretical knowledge with experimental data in a Physics-guided Machine Learning setting. A proposed three-system architecture comprised a model-based estimation, actual data-driven model training, and experimental validation. The proposed Physics-guided architecture has been implemented using simulated data and validated using experimental data from a Shock Dyno Suspension test rig. The developed algorithm draws its roots from a base-excitation suspension model and feeds upon the sprung and unsprung mass accelerations to control the damping characteristics of a semi-active suspension system in real-time. This control framework has been compared with the classical suspension control algorithms - Skyhook and Groundhook control based on the performance metrics of comfort cost about the chassis frequency zone.