Investigation of cycle-to-cycle variations in a spark-ignition engine based on a machine learning analysis of the early flame kernel

Abstract High-speed Mie scatter imaging and in-cylinder pressure measurements are performed in an optically accessible four-stroke spark-ignition engine to investigate cycle-to-cycle variations (CCVs). Droplet Mie scattering is used to measure the cross-sectional flame contour. Machine learning (ML) methods are applied to predict combustion cycles of high maximum in-cylinder pressure based on flame cross-sections at -15 °CA (crank angle degrees before top dead center). All tested ML methods (decision tree-based, multi-layer perceptron, and logistic regression) are able to predict high energy engine cycles given only partial flame topology information at a temporal snapshot of the flame propagation. Feature importance analyses, employed using decision tree methods, reveal that a combination of flame position and shape features of the flames' cross-section is necessary in achieving a high prediction accuracy. These sensitivity analyses can be used to gain insight into the combustion processes and to strategically reduce the number of features. This enables the addition of new physical quantities in future investigations.