Application of Extreme Learning Machine to Knock Probability Control of SI Combustion Engines

Abstract

In spark-ignition gasoline engines, spark timing is optimized for fuel economy and power output. Under some heavy load and low engine speed operating conditions, however, an increased spark advance can often lead to a frequent occurring of knock. As a compromise between the engine power output and the risk of knock, various spark timing controllers have been proposed to regulate the spark timing so that a low knock probability is tolerated for higher torque output. Since the binary knock signal contains little information about the change of the engine operating conditions, a feedforward map can speedup the responding speed of the controller. In this work, an ELM is used to learning the relation between the knock probability and the operating condition offline and online. The probability estimation of the ELM is then used to determine the initial spark timing of the likelihood ratio-based controller for the on-board knock probability control. The proposed control method is also validated on a full-scale engine test bench with a production engine.