Modelling of Gasoline Engine-Out Emissions Using Artificial Neural Networks

"Due to ever shorter time-to-market requirements and a simultaneous increase in powertrain complexity, the challenges in the field of gasoline powertrain calibration are growing. In addition, the great variety of vehicle variants requires an increasing number of prototype vehicles for calibration and validation tasks within the framework of the current Real Driving Emissions (RDE) regulations and the expected Euro 7 emission standards. Hardware-in-the-Loop (HiL) approaches have been introduced successfully to support the calibration tasks in parallel to the conventional vehicle development activities. Using highly sophisticated simulation models, the HiL approach enables a more reliable compliance with the emission limits and improves the quality of calibrations, while reducing the number of required prototype vehicles, test resources and thus overall development costs. To further improve the quality, this paper presents a novel real-time simulation model that aims to predict the exhaust emissions of a gasoline engine in virtual driving cycles to support the HiL-based virtual emission calibration process. This real-time emission model is based on Artificial Neural Networks (ANN) and can simulate the gaseous engine-out emissions during stationary operation of the engine as a function of all relevant quantitatively measurable and physically relevant influencing factors. To enable the emission simulation during real driving cycles, correction factors are added to the ANN to account for the influence of transient engine behavior on the engine-out emissions. In a first step, it is shown how 20,000 data points from engine test bench measurements are used to train the ANN. As result, a high reproduction quality is achieved, with a maximum deviation of the simulated gaseous engine raw emissions and the real vehicle measurements during driving cycle operation of 4 %. In a second step, the number of training data is reduced to demonstrate the influence of the total number of data points on the simulation accuracy. With a reduction of the number of training data points by more than 80 %, a simulation accuracy of about 4 % can be maintained."