Air-fuel ratio control of a gasoline engine by means of a recurrent neural network with differentiated input

Abstract

A neural controller for the air-fuel mixture in a gasoline engine has been developed. The catalyst requires the air-fuel ratio to be kept at the stoiquiometric value. Conventional systems are not able to avoid important excursions from the set point during transient operation.

First a mathematical model of the engine has been designed. It has been validated with experimental data, and a simple static feedforward controller plus a PID feedback controller.

An observer based on a neural network is used to close the loop instead of the lambda sensor, which enables the tuning of the observer. A recurrent neural network has been developed starting from the Elman network, and separating the context neurons in as many groups as the network has inputs. Each group is trained separately, thus adapted to the particular dynamics of the input with which it is associated.