On-board knock probability map learning–based spark advance control for combustion engines

Abstract

This article presents an on-board map learning–based spark advance control framework for combustion engines. The proposed control framework addresses the knock probabilistic constrained thermal efficiency optimization problem with three layers. First, in the upper layer, maps of knock event distribution and thermal efficiency are learned with manifold pressure and combustion phase as inputs. Second, the middle layer generates the knock probability constrained optimal combustion phase reference that is subsequently tracked by a hypothesis test-based feedback controller. Third, the lower layer employs a partial likelihood-based knock controller that retards the spark advance in case of the frequent knock events. The key contributions of this work are the three-layer control framework and the knock event distribution map learning in the upper layer. The knock event is supposed to obey binomial distribution, and the distribution is modeled by beta distribution and learned in the perspective of Bayesian learning. Moreover, the normalization algorithm is proposed for online feedfoward map update. The proposed map learning–based spark advance control framework is experimentally validated in a test bench equipped with a spark-ignition engine.