Development of a Machine Learning model to characterize the performance of a Selective Catalytic Reduction on Filter after-treatment system for a Heavy-Duty Diesel Engine

Abstract

Particulate matter (PM) and Oxides of Nitrogen (NOx) are the major pollutants in diesel engines, an attempt to control one leads to an increase in the other, a phenomenon known as PM-NOx trade-off in diesel engine emission control. Currently, these two pollutants are controlled by the Diesel Particulate Filter (DPF) and the Selective Catalytic Reduction (SCR) after-treatment system respectively, in addition to the Diesel Oxidation Catalyst (DOC) which helps to provide 1:1 split of NO/NO2 and helps with raising exhaust gas temperatures. Today, heavy-duty diesel engines feature a DPF, a primary SCR and a secondary SCR. Despite this complex after-treatment control system, current emissions are still well above the limits that have been set for 2024 and beyond. Emission regulations are continually getting more stringent, the California Air Resources Board (CARB) had proposed and adopted a low NOx emission regulation in the year 2020. This rule tightens Federal Test Procedure (FTP) NOx limits to 0.05 g/bhp-hr from 2024 and 0.02 g/bhp-hr from 2027. This is about 90 % reduction from the current 0.2 g/bhp-hr limit by 2027. The current after-treatment configuration for Heavy Duty Diesel Engines (HDDEs) is already too voluminous and costly. Integrating the SCR and the DPF into a single component known as the Selective Catalytic Reduction on Filter (SCRF) could reduce the cost and save some space as well as help improve the thermal inertia and reduce cold start emissions in HDDEs. However, NO2 is an important part of the operations of both the DPF (passive soot oxidation) and the SCR (fast SCR reactions), therefore, there could be a struggle for NO2 consumption in both DPF and SCR activities of the SCRF, this could limit the performance in the path/operation that is not being favored in the struggle. The main purpose of this study is to characterize the performance of an SCRF and determine if certain functionality is favored over another, as well as investigate the extent of the impact on the less favored functionality of the SCRF. The SCRF was loaded with soot from 0 g/L to 5 g/L and NOx conversion and soot oxidation was examined as soot loading increased on an engine dynamometer test. A data-driven model was also developed to predict NOx conversion efficiency of the SCRF and investigate the possibility of the model being utilized as a virtual NOx sensor to save cost on expensive NOx sensor at post SCRF location. Results