Weighting of Parameters in Artificial Neural Network Prediction of Heavy-Duty Diesel Engine Emissions 2002-01-2878
The use of Artificial Neural Networks (ANNs) as a predictive tool has been shown to have a broad range of applications. Earlier work by the authors using ANN models to predict carbon dioxide (CO2), carbon monoxide (CO), oxides of nitrogen (NOx), and particulate matter (PM) from heavy-duty diesel engines and vehicles yielded marginal to excellent results. These ANN models can be a useful tool in inventory prediction, hybrid vehicle design optimization, and incorporated into a feedback loop of an on-board, active fuel injection management system. In this research, the ANN models were trained on continuous engine and emissions data. The engine data were used as inputs to the ANN models and consisted of engine speed, torque, and their respective first and second derivatives over a one, five, and ten second time range. The continuous emissions data were the desired output that the ANN models learned to predict through an iterative training process. The ANN models provide a valuable insight into the effect of different engine parameters on emissions constituents. The normalized weighted input values derived by the ANN during the iterative training process reflects the relative importance of the variables. Artificial neural network models were developed for CO2, CO, NOx, and PM emissions from a 1999 Cummins ISM 370 engine. The Cummins engine was exercised through the Federal Test Procedure (FTP), the European Stationary Cycle (ESC), the European Transient Cycle (ETC or FIGE), and a Random Cycle Generator (RCG). The weighted model inputs showed (as expected) speed and torque to be the dominant engine parameters in the formation of CO2 and NOx, and the transient engine parameters to be more significant in the formation of CO and PM. Comparison of the weighted inputs from the different cycles provided further insight into the effects of engine test cycles on emissions.