A Virtual Sensor for Predicting Diesel Engine Emissions from Cylinder Pressure Data

Cylinder pressure sensors provide detailed information on the diesel engine combustion process. This paper presents a method to use cylinder-pressure data for prediction of engine emissions by exploiting data-mining techniques. The proposed method uses principal component analysis to reduce the dimension of the cylinder-pressure data, and a neural network to model the nonlinear relationship between the cylinder pressure and emissions. An algorithm is presented for training the neural network to predict cylinder-individual emissions even though the training data only provides cylinder-averaged target data. The algorithm was applied to an experimental data set from a six-cylinder heavy-duty engine, and it is verified that trends in emissions during transient engine operation are captured successfully by the model.