A Model-less Approach for Estimating Vehicles Sideslip Angle by a Neural Network Concept

Abstract

Over the past few years, vehicle dynamics systems have been constantly improved by new technologies due to the rapid advance in computational systems and, so, have been continually developed to enhance vehicle handling and safety of the passengers. Unfortunately, these systems require information of variables not always easy to be accessed without proper sensing. In this way, several studies had been carried out to obtain a good estimation of the variables of interest as tire-ground interaction forces and sideslip angle, to mention a few. Also, the sideslip angle relates directly to the vehicle's lateral dynamics, an important factor for control stability. Most of these studies are based on vehicles dynamics and tire models, but non-linearities contribute to the low accuracy of the models. The present work performed several simulation tests using a well-known software to access vehicle data, which was used to train a machine learning technique, neural network, to predict the vehicle body sideslip angle along the trajectory. Such tests considered high speed profiles and different friction coefficient values in order to reach as closely the slip limit. Such tests were studied to indicate which set of maneuvers could be better to obtain sideslip information within the operational range of the vehicle. The best result was obtained when training a neural network with information from slalom and lane change tests, with a validation error lower than a 1%. Although, other set of maneuvers needs to be included to the training data so the model could cover a wider range of driving events.