Neural Networks Applied to Ignition Timing Calibration

Abstract

The present control methods for combustion parameters in engine management systems, such as ignition timing or desired air-to-fuel ratio, are based on “look-up” tables. Optimised engine parameters are accessed only for specific input values. In the intermediate points the output parameters are linearly interpolated, which results in the engine running in sub-optimal conditions. This thesis discusses the feasibility of replacing these look-up tables with non-linear mappings produced by artificial neural networks. The thesis reports the experiments for two data sets collected from two Rover internal combustion engines. The preliminary experiments carried out for the air-to-fuel ratio and the ignition timing data show that non-linear models, such as the Radial Basis Function Networks, Multilayer Perceptron and the committees of these networks, can be used to produce smooth and accurate mappings of the engine parameters. The neural network approach is feasible and provides a new and more efficient way to handle the problem of controlling the engine parameters. The thesis also reports on the C++ neural network library created for use in this project.