A Structure and Calibration Method for Data-Driven Modeling of NOX and Soot Emissions from a Diesel Engine

Journal article, 2012

The development and implementation of a new structure for data-driven models for NOX and soot emissions is described. The model structure is a linear regression model, where physically relevant input signals are used as regressors, and all the regression parameters are defined as grid-maps in the engine speed/injected fuel domain. The method of using grid-maps in the engine speed/injected fuel domain for all the regression parameters enables the models to be valid for changes in physical parameters that affect the emissions, without having to include these parameters as input signals to the models. This is possible for parameters that are dependent only on the engine speed and the amount of injected fuel. This means that models can handle changes for different parameters in the complete working range of the engine, without having to include all signals that actually effect the emissions into the models. The approach possibly also enables for the model to handle the main differences between steady-state engine operation and transient engine operation, thus possibly being able to use steady-state engine measurement data to calibrate the model, but still achieve acceptable performance for transient engine operation. This, however, is not evaluated in this study. The model structure has been used to create models for NOX and soot emissions. These models have been calibrated using measured steady-data from a 5 cylinder Volvo passenger car diesel engine with a displacement volume of 2.4 liters, equipped with a turbocharger, an exhaust gas recirculation system, and a common rail injection system. The models estimate NOX mass flow with a root mean square error of 0.0021 g/s and soot mass flow with a root mean square error of 0.59 mg/s for the steady-state engine data used in this study. The models are capable of reacting to different calibratable engine parameters, and they are also fast to execute. This makes them suitable for development of engine management system optimization. The models could also be implemented directly into an engine management system. For comparison, three other fast models of different types for NOX and soot emissions have been implemented and evaluated.