A Machine Learning Approach for Engine Model-Based Control on NOx Emissions

Increasingly stringent regulations on engine emissions require strict control of nitrogen oxide (NOx) emissions in diesel engines. Feedback control systems coupled with virtual sensors for real-time NOx readings have shown to be effective solutions for managing emissions. The authors of this paper propose a machine learning approach for developing a virtual NOx sensor implemented on an Engine Control Unit (ECU) of a YANMAR diesel engine. A Random Forest model was trained on data comprising Ramped Modal Cycles (RMCs) and Non-Road Transient Cycles (NRTCs) with a focus on robustness with respect to engine-to-engine variability in ECU sensor reading. Despite strong constraints imposed on the complexity of the model due to the limited computing power of the ECU, good prediction performance was obtained on both cycles (R² = 1.0 on RMCs and R² = 0.967 on NRTCs). The present study shows that machine learning models trained on transient data can play an important role in developing robust NOx emissions control systems on diesel engines.