Development of a Machine-Learning Classification Model for an Electrochemical Nitrogen Oxides Sensor in Gasoline Powertrains

Abstract

Future automotive emission regulations are becoming increasingly dependent on off-cycle (acquired on road and referred to as “real-world”) driving and testing. This was driven in part by the often-observed fact that laboratory emission drive cycles (developed to evaluate a vehicle’s emissions on a chassis dynamometer) may not fully capture the nature of real-world driving. As a result, portable emission measurement systems were developed that could be fit in the trunk of a vehicle, but were relatively large, expensive, and complex to operate. It would be advantageous to have low-cost and simple to operate on-board sensors that could be used in a gasoline powertrain to monitor important criteria emission species, such as NOx. The electrochemical NOx sensor is often used for emissions control systems in diesel powertrains and a proven technology for application to the relatively harsh environment of automotive exhaust. However, electrochemical NOx sensors are nearly equally sensitive to both NOx and NH3, setting up an implicit classification problem that must be solved before they can accurately measure NOx. In this work, we develop a machine-learning model to classify the output of a NOx sensor in a gasoline powertrain. A model generalization study is conducted, and the model is found to be ~96% accurate and able to predict NOx mass emitted over a drive cycle within ~9% of a perfectly classified NOx sensor.