Development of a cold-start emission model for diesel vehicles using an artificial neural network trained with real-world driving data.

During the cold start and warm-up phase, modern vehicles emit considerable amounts of pollutants due to the incomplete combustion and deteriorated performance of aftertreatment devices. In terms of emission modeling, there have been many attempts to estimate cold start emission such as cold-hot conversion factor, regression model, and physis-based model. However, as the emission characteristic become complicated due to the adoption of aftertreatment devices and various emission control strategies for the strengthened emission regulations, the conventional cold start emission models do not always show satisfactory performances. In this study, artificial neural networks were used to predict the cold start emissions of carbon dioxide, nitrogen oxides, carbon monoxide, and total hydrocarbon of diesel passenger vehicles. We used real-world driving data to train neural networks as an emission prediction tool. Through machine leaning, numerous trainable variables of neural networks were properly adjusted to predict cold start emissions. For input variables of the ANN model, the velocity, vehicle specific power, engine speed, engine torque, and engine coolant temperature were used. The proposed ANN models accurately predicted sharp increases in carbon monoxide, hydrocarbon, and nitrogen oxides during the cold start phase. In addition to the quantitative estimations, the correlations between the operating variables and exhaust gas emissions were visually described in the form of emission maps. The emission map graphically showed the emission levels according to the vehicle and engine operating parameters.