Zhai, Y (2009) Advanced Neural Network Based Control for Automotive Engines. Doctoral thesis, Liverpool John Moores University.

Preview

Text 515359.pdf - Published Version Download (6MB) | Preview

Abstract

This thesis investigates the application of artificial neural networks (NN) in air/fuel ratio (AFR) control of spark ignition(SI) engines. Three advanced neural network based control schemes are proposed: radial basis function(RBF) neural network based feedforward-feedback control scheme, RBF based model predictive control scheme, and diagonal recurrent neural network (DRNN) - based model predictive control scheme. The major objective of these control schemes is to maintain the air/fuel ratio at the stoichiometric value of 14.7 , under varying disturbance and system uncertainty. All the developed methods have been assessed using an engine simulation model built based on a widely used engine model benchmark, mean value engine model (MVEM). Satisfactory control performance in terms of effective regulation and robustness to disturbance and system component change have been achieved. In the feedforward-feedback control scheme, a neural network model is used to predict air mass flow from system measurements. Then, the injected fuel is estimated by an inverse NN controller. The simulation results have shown that much improved control performance has been achieved compared with conventional PID control in both transient and steady-state response. A nonlinear model predictive control is developed for AFR control in this re- . search using RBF model. A one-dimensional optimization method, the secant method is employed to obtain optimal control variable in the MPC scheme, so that the computation load and consequently the computation time is greatly reduced. This feature significantly enhances the applicability of the MPC to industrial systems with fast dynamics. Moreover, the RBF model is on-line adapted to model engine time-varying dynamics and parameter uncertainty. As such, the developed control scheme is more robust and this is approved in the evaluation. The MPC strategy is further developed with the RBF model replaced by a DRNN model. The DRNN has structure including a information-storing neurons and is therefore more appropriate for dynamics system modelling than the RBF, a static network. In this research, the dynamic back-propagation algorithm (DBP) is adopted to train the DRNN and is realized by automatic differentiation (AD) technique. This greatly reduces the computation load and time in the model training. The MPC using the DRNN model is found in the simulation evaluation having better control performance than the RBF -based model predictive control. The main contribution of this research lies in the following aspects. A neural network based feedforward-feedback control scheme is developed for AFR of SI engines, which is performed better than traditional look-up table with PI control method. This new method needs moderate computation and therefore has strong potential to be applied in production engines in automotive industry. Furthermore, two adaptive neural network models, a RBF model and a DRNN model, are developed for engine and incorporated into the MPC scheme. Such developed two MPC schemes are proved by simulations having advanced features of low computation load, better regulation performance in both transient and steady state, and stronger robustness to engine time-varying dynamics and parameter uncertainty. Finally, the developed schemes are considered to suit the limited hardware capacity of engine control and have feasibility and strong potential to be practically implemented in the production engines.

Item Type:	Thesis (Doctoral)
Subjects:	Q Science > QA Mathematics > QA76 Computer software T Technology > TL Motor vehicles. Aeronautics. Astronautics
Divisions:	Engineering
Date Deposited:	14 Mar 2017 12:14
Last Modified:	03 Sep 2021 23:30
DOI or ID number:	10.24377/LJMU.t.00005933
URI:	https://researchonline.ljmu.ac.uk/id/eprint/5933

View Item