Spark ignition engine port air mass flow prediction using artificial neural networks

In order to maintain the air fuel ratio within the stoichiometric operating window, which is necessary for efficient catalytic converter operation an accurate estimation of the mass airflow at the engine ports is critical. It is difficult to accurately represent the port air mass flow of an Internal Combustion Spark Ignition engine as reciprocating engines are complex non-linear systems based on a large number of interrelated parameters. Conventional air fuel ratio control strategies use a number of three dimensional feedforward look up tables to represent these complex nonlinear engine functions. These look up tables are usually functions of only two engine variables, engine speed and engine load. Engine load is either, calculated from the speed density relationship using an absolute manifold air pressure sensor or measured directly using a mass air flow sensor. Conventional AFR control algorithms perform poorly during transient conditions as the strategies inherent in look up tables, are based on stationary or non-dynamic modelling techniques. Modern air fuel ratio control strategies employ a large number of correction factors to compensate for transient engine operation. This research is a preliminary investigation into the feasibility of using Artificial Neural Networks to represent transient nonlinear engine functions. This research develops offline artificial neural network models of the port air mass flow of a Spark Ignition, based on Hendricks' et al [1] accepted mean value engine model description of the manifold filling phenomena. In particular two different Artificial Neural Network paradigms, namely the Backpropagation algorithm and the fast converging Optimise Layer by Layer algorithm will be trained on data collected through both steady state and transient chassis dynamometer testing. Both steady state and transient air mass flow models will be developed in this investigation. The accuracy of the Backpropagation and Optimised Layer by Layer models will be analysed both qualitatively and quantitatively and compared in terms of the Root Mean Squared percentage error and computational time in an effort to evaluate the most appropriate model for future online engine implementation.