The use of alternative fuels, such as biofuels and synthetic fuels in small mobility engines has become more common these days. Although these fuels contribute to the carbon neutrality, it is known that they do not have a certain fuel composition, which significantly affects the combustion characteristics of an engine, such as knocking and combustion duration. Therefore, to get the most out of these sustainable fuels, it is necessary to develop engine systems that are highly robust to variations in fuel composition. To achieve this goal, a method to sense fuel characteristics onboard using sensors already widespread in use or can be installed inexpensively is required. Although in-cylinder piezoelectric pressure sensors are useful for research in the laboratory, it is not suitable for the use in commercial engines because of its high cost. Therefore, the use of other sensors should be considered. The purpose of this study is to experimentally analyze what information related to combustion and fuel can be obtained from multiple cost-effective sensors mounted on an engine. For that goal, a linear multiple regression model and Neural Network (NN) model was developed to estimate fuel’s Lower Heating Value (LHV) and combustion duration.
Experiments were conducted on a 4-cylinder spark ignition (SI) engine, and combustion characteristics of multiple fuels were investigated while varying engine operating conditions. In addition to gasoline, CH4 gas was introduced into cylinders to simulate the change in fuel composition. Sensors used in this study include intake and exhaust pressure sensors, thermocouples, and in-cylinder ion current sensor. Selection of input variables (sensors) for the regression models were done based on the results of the experiment, and linear multiple regression model and NN model were developed. The prediction errors (RMSE) for LHV were 0.54 MJ/kg with linear regression model and 0.95 MJ/kg with NN model. For CA10-90, prediction errors were 6.15 deg with linear regression model and 14.48 deg for NN model. Since the accuracy of the models were not high enough, hyperparameter tuning was done using Bayesian optimization, and prediction accuracies were improved. However, further work, such as building physical model,increasing sample size, or adding extra sensors, must be done to use these models for engine control.