Machine Learning-based Classification of Combustion Events in an RCCI Engine Using Heat Release Rate Shapes

Abstract

Reactivity controlled compression ignition (RCCI) mode offers high thermal efficiency and low nitrogen oxides (NOx) and soot emissions. However, high cyclic variability at low engine load and high pressure rise rates at high loads limit RCCI operation. Therefore, it is important to control the combustion event in an RCCI engines to prevent abnormal engine combustion. To this end, combustion in RCCI mode was studied by analyzing the heat release rates calculated from the in-cylinder pressure data at 798 different operating conditions. Five distinct heat release shapes are identified. These different heat release traces were characterized based on start of combustion, burn duration, combustion phasing, maximum pressure rise rate, maximum amount of heat release, maximum in-cylinder gas temperature and pressure. Both supervised and unsupervised machine learning approaches are used to classify different types of heat release rates. K-means clustering, an unsupervised algorithm, could not cluster the heat release traces distinctly. Convolution neural network (CNN) and decision trees, supervised classification algorithms, were designed to classify the heat release rates. The CNN algorithm showed 70% accuracy in predicting the shapes of heat release rates while decision tree resulted in 74.5% accuracy in predicting different heat release rate traces.