Model structure selection in powertrain calibration and control



Li, Zongyan
Model structure selection in powertrain calibration and control. Doctor of Philosophy thesis, University of Liverpool.

[img] PDF (PhD Thesis)
LiZongyan_October2012_12013.pdf - Author Accepted Manuscript
Available under License Creative Commons Attribution No Derivatives.

Download (2MB)

Abstract

This thesis develops and investigates the application of novel identification and structure identification techniques for I.C. engine systems. The legislated demand for reduced vehicle fuel consumption and emissions indicates that improved model-based dynamical engine calibration and control methods are required in place of the existing static set-point based mapping methods currently used in industry. The choice of structure of any dynamical engine model has significant consequences for the accuracy and the calibration/optimization time of engine systems. This thesis primarily addresses the issue of this structure selection. Linear models are well understood and relatively easy to implement however the modern I.C. engine is a highly nonlinear system which restricts the use of linear structures. Further the newer technologies required to achieve demanding fuel consumption and emission targets are increasingly more complex and nonlinear. The selection of appropriate nonlinear model regressor terms presents a combinatorial explosion problem which must be solved for accurate engine system modelling. In this thesis, two systematic nonlinear model structure selection techniques, namely stepwise regression with F-statistics and orthogonal least squares method with error reduction ratio, are accordingly investigated. SISO algebraic NARMAX engine models are then established in simulation studies with these methods and demonstrate the effectiveness of the approach. The thesis also investigates the development and application of multi-modelling techniques and the expansion of the model structure selection techniques to the identification of the local models terms within the multi-model structures for the engine. Based on the en- gine operating regions, novel multi-model networks can be established and several alternative multi-modelling techniques, such as LOLIMOT, Neural Network, Gaussian and log-sigmoid function weighted multi-models, for the multi-model engine system identification are explored and compared. An experimental validation of the methods is given by a black box identification of SISO engine models which are developed purely from the experimental engine test data sets. The results demonstrate that the multi-model structure selection techniques can be successfully applied on the engine systems, and that the multi-modelling techniques give good model accuracy and that good modelling efficiency can also be achieved. The outcome is a set of techniques for the efficient development of accurate nonlinear black-box models which can be acquired from experimental dynamometer test-bed data which should assist in the dynamic control of future advanced technology engine systems.

Item Type: Thesis (Doctor of Philosophy)
Additional Information: Date: 2012-10 (completed)
Divisions: Faculty of Science and Engineering > School of Engineering
Depositing User: Symplectic Admin
Date Deposited: 28 Aug 2013 10:40
Last Modified: 16 Dec 2022 04:39
DOI: 10.17638/00012013
Supervisors:
URI: https://livrepository.liverpool.ac.uk/id/eprint/12013