Advanced Neural Network Engine Control

The main subject of this dissertation is the dynamic modelling of Spark Ignition
(SI) engines. This is done on the level of Mean Value Engine Models (MVEMs).
This modelling is done using physical modelling techniques as well as using neural networks. One of the main aims of the dissertation is to compare these two modelling techniques with respect to accuracy. Engine subsystems are modelled by dynamic neural networks trained with the predictive neural network training algorithm developed in this work and the results are compared with the physical Adiabatic Mean Value Engine model’s corresponding signals.

In addition, neural networks’ accuracy as virtual sensors for O₂ concentration in the exhaust, in-cylinder air fuel ratio, in-cylinder peak pressure and peak pressure location is tested for a 2.0 L Puma diesel engine. A predictive cost function based neural network training algorithm for improved dynamic neural network model training has been developed. Furthermore, the necessary mathematics for a fast C++ implementation as well as a C++ implementation of a matrix library and the predictive neural network training algorithm has been developed.
A general neural network based multi input multi output predictive controller algorithm has been developed and the necessary mathematics for a fast C++ implementation is derived. A C++ implementation of the neural network based predictive controller has been developed. A proof of stability is given for the predictive control strategy utilized if a final state cost is added to the cost function.