Real-Time Capable and Modular Modeling of Wheel Suspensions using Neural Networks

The paper describes the development of a real-time capable simulation model for a vehicle suspension using Neural Networks (NN). In addition to the simulation of the dynamics, the model also allows the real-time computation of the reaction forces acting in the joints. First, a highly accurate, object-oriented model is created, which provides the necessary training, test and validation data for the neural network. The method delivers close-to-reality results because the elasto-kinematic behavior of the system is also considered. Secondly, an approach for NN is developed, which is based on the NARX model (Non-linear Auto Regressive eXogenous). Both, the model for generating the data and the NN are implemented in the MATLAB/SIMULINK software to ensure a high degree of compatibility. The generalization ability is demonstrated in several different driving maneuvers. It is shown that the NN can effectively model wheel suspension accelerations and bearing forces and requires significantly less computing time than a DAE-based (Differential Algebraic Equations) simulation.