Tailpipe NOx Emissions Modeling of a Heavy-Duty Diesel Truck Using Deep Learning Methods

This study develops deep learning (DL) long–short-term memory (LSTM) models to predict tailpipe nitrogen oxides (NOx) emissions using real-driving on-road data from a heavy-duty Class 8 truck. The dataset comprises over 4 million data points collected across 11,000 km of driving under diverse road, weather, and load conditions. The effects of dataset size, model complexity, and input feature set on model performance are investigated, with the largest training dataset containing around 3.5 million data points and the most complex model consisting of over 0.5 million parameters. Results show that a large and diverse training dataset is essential for achieving accurate prediction of both instantaneous and cumulative NOx emissions. Increasing model complexity only enhances model performance to a certain extent, depending on the size of the training dataset. The best-performing model developed in this study achieves an R² higher than 0.9 for instantaneous NOx emissions and less than a 2% error for cumulative NOx emissions on the test data. Furthermore, the model achieves an F1 score above 0.9 in determining whether NOx emissions comply with emission standards. The developed DL tailpipe emission models in this study have diverse applications based on the amount and type of available input data, including engine and aftertreatment system control, diagnostics, and vehicle system-level simulations. These applications collectively contribute to minimizing NOx emissions of vehicles to meet stringent transportation emission standards.