Parameter estimation for lateral dynamic tire models using empirical inertial data

A comparative case study of gradient-based algorithms vs genetic algorithms

Master thesis (2018)

Authors

D.H. den Hartog Mechanical Engineering

Contributors

R. Happee (supervisor 1)
B. Shyrokau (supervisor 1)
Faculty
Mechanical Engineering
Copyright: © 2018 Daniel den Hartog
Machine Learning Optimization Algorithm Parameter Estimation Vehicle dynamics Magic Formula Tire Dynamics
More Info
expand_more

Published Date

02-05-2018

Language

English

Reuse Rights

Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.

Faculty

Mechanical Engineering

Abstract

Self-driving vehicles are the future of automotive engineering. Systems that take over control from the driver are developed to be able to interact with the conditions of the road and other obstacles. To develop these systems, developers use vehicle models to simulate the behaviour of the moving vehicle. The systems developed using these models are transferred the vehicle and are expected to perform accordingly. Therefore it is important that the vehicle models reflect the behaviour of the actual vehicle.

This thesis investigates the methods used to estimate the model parameters which influence the resulting vehicle simulation. Experimental tests were carried out to acquire vehicle body inertial measurement data. This data was used to minimize lateral acceleration and yaw-rate error produced by the output of a single track vehicle model using a simplified Magic formula developed by H. Pacejka.
Deterministic gradient-based algorithms, such as multi-objective Sequential Quadratic Programming (SQP), are frequently used as a way to optimize an initial approximation of the model parameter set. It is shown that this initial parameter strongly influences results of the algorithm due to local minima.

As an alternative, genetic algorithms were investigated to minimize the influence of the initial parameter set. An adaptive component, Temporal Difference Q Learning (TDQL), was also added to further reduce the input of the user and to increase the performance of the algorithm.

Four algorithms, of various complexity, were implemented in Matlab and executed. The performance of the resulting parameter sets, as well as the performance of the algorithms themselves, were analysed and compared. It is concluded in this thesis that the adaptive genetic algorithm performs slightly better than a simple gradient-based algorithm with respect to the objectives and the duration of the algorithm. However, a genetic algorithm without TDQL is recommended for its good performance and the simulation flexibility of the results for this simple vehicle and tire model.