Evaluating the feasibility of machine learning algorithms for combustion regime classification in biodiesel-fueled homogeneous charge compression ignition engines

Abstract

The present study explores the feasibility of machine-learning algorithms in classifying the combustion regimes in a homogenous charge compression ignition (HCCI) engine fueled with biodiesels from diverse sources. A distinctive feature of this research is the extensive database of HCCI combustion regimesgenerated for neat biodiesel fuels of different compositions. The engine experiments reveal significant variability in the load range of operation of HCCI engines across different biodiesel fuels. Strategies including compression ratio variation and charge dilution were employed to achieve higher engine loads, and the results confirm that the interplay of fuel reactivity and engine load, along with compression ratio and charge dilution, dictates the combustion stability. This complex phenomenon necessitates ascertaining the prevailing combustion type under a given engine load for biodiesel fuels derived from diverse sources. Thus, the present study aims to investigate the applicability of machine learning algorithms for developing models that classify combustion regimes in a neat biodiesel-fueled light-duty HCCI diesel engine. Models are developed to categorize biodiesel HCCI combustion into four groups: engine misfire, stable combustion without dilution, stable combustion with dilution, and knocking combustion. Biodiesel composition (13 different methyl esters), compression ratio and engine load are considered the input variables for the machine learning (ML) models. The results show that Naive Bayes and logistic regression models exhibit poor applicability among the investigated ML models due to a low F1 score (<0.8). The k-NN and SVM models demonstrate practical suitability with reasonably good F1 scores. However, the decision tree model outperforms all, precisely classifying 94 % of calibration and validation samples. It could precisely classify all the validation samples corresponding to misfire combustion, 21 of 22 samples to stable combustion, 4 of 6 samples to stable combustion with dilution, and 30 of 31 to knocking combustion. Thus, it emerges as the best approach to classify the combustion regimes of HCCI engines operated with neat biodiesel fuels.