Machine-Learning-Based Emission Models in Gasoline Powertrains—Part 2: Virtual Carbon Monoxide

In this work, tailpipe carbon monoxide emission from a gasoline powertrain case study vehicle was analyzed for off-cycle (i.e., on road) driving to develop a virtual sensor. The vehicle was equipped with a portable emissions measurement system (PEMS) that measured carbon monoxide concentration and exhaust volumetric flowrate to calculate the mass of carbon monoxide emitted from the tailpipe. The vehicle was also equipped with a tailpipe electrochemical NO_x sensor, and a correlation between its linear oxygen signal and the PEMS-measured carbon monoxide concentration was observed. The NO_x sensor linear oxygen signal depends on the concentration of several reducing species, and a machine learning model was trained using this data and other features to target the PEMS-measured carbon monoxide mass emission. The model demonstrated a mean absolute percentage error (MAPE) of 19% when using 15 training drive cycles. Finally, a virtual carbon monoxide sensor was developed by removing the tailpipe NO_x sensor information from the model feature set and predicting tailpipe carbon monoxide mass. The virtual model MAPE was shown to increase by 5% compared to the earlier version with a tailpipe NO_x sensor over the same number of training, validation, and test drive cycles. The minimal degradation in accuracy for the virtual model was hypothesized to result from the fact that narrowband oxygen sensors may contain information regarding how rich or lean the exhaust gas is compared to stoichiometric conditions. This is analogous to the information provided by a wide-band oxygen sensor, but potentially with reduced resolution and accuracy. The data-driven approach was able to produce a novel virtual tailpipe carbon monoxide sensor in a gasoline powertrain using only common powertrain and emission sensors.