A Neural Network Implementation of Peak Pressure Position Control by Ionization Current Feedback

Abstract

This paper describes a neural-network (NN)-based scheme for the control of a cylinder peak pressure position (PPP)—also known as the location of peak pressure (LPP)—by spark timing in a gasoline internal combustion engine. The scheme uses the ionization current to act as a virtual sensor, which is subsequently used for PPP control. A NN is trained offline on principal-component analysis data to predict the cylinder peak pressure position under dynamically varying engine load, speed, and spark advance (SA) settings. Experimental results demonstrate that the PPP prediction by the NN correlates well with those measured from in-cylinder pressure sensors across transients of load, SA, and engine speeds. The dynamic training data allow rapid model identification across the identified engine range, as opposed to just fixed operating points. A linear robust constrained-variance controller, which is a robustified form of the minimum variance controller, is used to regulate the PPP by SA control action, using the NN as a PPP sensor. The control scheme is validated by experimental implementation on a port fuel-injected four-cylinder 1.6 l gasoline internal combustion engine.