A grey-box machine learning based model of an electrochemical gas sensor

Abstract

A grey-box machine learning based model of an electrochemical O2–NOx sensor is developed using the physical understanding of the sensor working principles and a state-of-the-art machine learning technique: support vector machine (SVM). The model is used to predict the sensor response at a wide range of sensor operating conditions in the presence of different concentrations of NOx and ammonia. To prepare a comprehensive training and test data set, the production sensor is first mounted on the exhaust system of a spark ignition, a diesel engine, and then on a fully controlled sensor test rig. The sensor is not modified, rather the sensor working temperature, all of the sensor cell potentials, and the pumping current of the O2 sensing cell are the model inputs that can be varied while the pumping current of the NOx sensing cell is considered as the model output. A 9-feature low order model (LOM) and a 45-feature high order model (HOM) are developed with linear and Gaussian kernels. The model performance and generalizability are then verified by conducting input-output trend analysis. The LOM with Gaussian kernel and the HOM with linear kernel have shown the highest accuracy and the best response trend prediction.