Predicting Tailpipe NOx Emission using Supervised Learning Algorithms

The dynamics underlying the tailpipe NOx emissions in vehicles is in a complex relationship with multiple engine components in a temporal manner. Physical models constructed to explain the relationship are too complex and solutions require high computational power. Additional uncontrollable environmental effects are also unable to incorporate these models. We present novel data-driven methods to predict tailpipe NOx emissions using machine learning and deep learning methods. As an ensemble methods representor eXtreme Gradient Boosting (XGBoost), as regularized linear models Elastic Net and as a deep learning method Long-Short-Term-Memory (LSTM) were chosen and their capability of handling time-series sensor data were assessed. We use engine speed, engine torque, pedal position and vehicle speed obtained from real-world driving data to deduce high NOx emission amount during real world driving. We reveal that satisfactory results can be achieved even if we have small data both in terms volume and variety by using weighted objective functions.