Investigation of Emission Characteristics during Low Temperature Combustion using Multivariate Adaptive Regression Splines.

摘要

Exhaust emissions from diesel engines operating in a low temperature combustion (LTC) regime are significantly affected by fuel composition and injection strategy. The starting point of this study is a collection of data correlating injection system parameters, and fuel characteristics, to response parameters such as engine-out emissions (oxides of nitrogen (NOx), total particulate matter (TPM), carbon monoxide (CO), hydrocarbons (HC)) and brake thermal efficiency (BTE).;The purpose of this work is to develop a statistical analysis tool to assist the emission analyst in modeling problems in which a response of interest is influenced by several variables and the objective is to optimize this response. The experimental data produced during LTC operation have been analyzed using an approach commonly known as Response Surface Methodology (RSM). Since the system under study may be responding to hidden inputs that are neither measured nor controlled, regression analysis must be performed via a flexible procedure. The methodology that will be used in this sense is called Multivariate Adaptive Regression Splines (MARS), which allows to approximate functions of many input variables given the value of the function at a collection of point in the input space.;Data was collected at West Virginia University's Engine and Emissions Research Laboratory for the project CRC AVFL-16. The test engine was a turbo-charged GM 1.9L operated in the LTC mode utilizing a split injection strategy. Main and pilot SOI timing and fuel split were varied per a 5 X 3 X 3 full factorial design. Advanced Vehicle Fuel Lubricants (AVFL) Committee of the Coordinating Research Council (CRC) defined a matrix of nine test Fuels for Advanced Combustion Engines (FACE) based on the variation of three properties: cetane number, aromatic content, and 90 percent distillation temperature. The experimental data was used has a platform for the code development, and for its validation.;Using multivariate data analysis is not only useful in visualizing correlations that otherwise would be hidden by the large amount of experimental data points, but it is also capable to predict the behavior of those points inside the domain where no data are available. As suggested by the name this is a regression methodology capable of adapting the shape of the regression splines to the data analyzed. Validation datasets which were independent of the `calibration' datasets were used to check the accuracy of the model predictions.