Machine learning-based analysis on cyclic variation of gasoline engine in stoichiometry

Abstract

Environmental problems have been getting serious at a terrific speed, and enhancement of thermal efficiency has been expected on engines. However, cyclic variation, a phenomenon whose output varies at each cycle in spite of steady operation, interferes with the enhancement of thermal efficiency. Not much has been revealed about the mechanism of cyclic variation and no specific ways of controlling it have been found. In this study, a neural network model with polynomial regression was constructed in order to try clarifying the mechanism of cyclic variation. Polynomial regression helps to visualize how much influence each input parameter has on a target parameter. In addition, the model was designed to eliminate unnecessary parameters in the regression equation all at once to enhance readability of the regression equation. Experiments were conducted on a 4-cylinder gasoline engine under stoichiometric conditions. Main data of those experiments were cylinder pressure, intake pressure and exhaust pressure of each cylinder. As input variables, these data are used in learning of the neural network model, and we get a regression equation, which has IMEP as an objective variable and input variables mentioned above as explanatory variables, as a model output. These input variables are the candidates which can be causes of cyclic variation. Analytical result showed that variables from other cylinder could affect cyclic variation of the target cylinder. Precision of the model was still quite low, but the model demonstrated its potential to show such a beneficial result. Therefore, it is necessary to improve the model precision. Considering other ways of selecting input variables may improve it.