Data-Based On-Board Diagnostics for Diesel-Engine Nox-Reduction Aftertreatment Systems

Atharva Tandale, Purdue University

Abstract

An aftertreatment (AT) system is responsible for reducing harmful exhaust emissions that result from fuel combustion in an internal combustion engine (ICE). Selective Catalytic Reduction (SCR) and Ammonia Slip Catalyst (ASC) are important components of diesel AT system that deal with oxides of nitrogen (NOx) and ammonia (NH3) components of the exhaust gas. SCR reduces NOx to harmless N2 and H2O by controlled injection of NH3 into the exhaust. NH3 itself being harmful, the unreacted NH3 that slips downstream of SCR is oxidized in ASC to N2. The performance of SCR-ASC system degrades with time as the SCR-ASC catalytic block ages. This leads to increase in emissions that eventually exceed the acceptable limits. On-Board Diagnostics (OBD) is a program that monitors this phenomenon in real time and flags when SCR-ASC system degrades beyond acceptable limit. The aim of this study is to devise a data-based OBD binary classification method to distinguish an aged SCR-ASC catalyst from healthy ones in real-world conditions. The data was made available by Cummins Inc. Available data consisted of various temporal signals such as Temperatures, Species Concentrations, Flow rates, etc, and was acquired in two different settings. The first was under Test-Cell (TC) settings, with controlled environment and lab-grade accurate sensors. The AT system was subjected to standard drive cycles, namely, cold Federal Test Procedure (cFTP), hot Federal Test Procedure (hFTP) and Random Mode Cycle (RMC). The second was under real-world operating conditions using commercial on-road truck sensors on 4 trucks at two different instances in time: one, during the truck’s on-road debut (referred to as DG trucks), and the second, after a significant number of miles on respective odometer readings (referred to as EUL trucks). In pure data-driven approach, all 5 measurable or computable signals namely, CE, T, F, DEF and EONOx, were used as input features, and class labels "1 " and "2 " were used for "DG" and "EUL" catalysts, respectively, as output feature. Multiple supervised Machine Learning models were cross-validated on TC data from all three cycles (cFTP, hFTP and RMC). Highest training-test accuracies were observed while using SVM with Gaussian kernel function model. Thus, an SVM with Gaussian kernel function model was used in subsequent analysis and is referred to as "classifier" from hereon. EUL truck with lowest mileage was deemed least aged and the one with most mileage was deemed most aged by the classifier cross-validated on filtered TC data. The classifier cross-validated on truck data showed a clear increase in catalyst age with mileage on all 4 trucks. In model-informed approach, TPNOx and storage fraction signals computed by a calibrated 3-state CSTR model and a look-up-table-based ASC map (referred to as "model" from hereon), were used in place of sensor measured signals. A classification TPNOx boundary was computed and a confidence interval was defined depending on position of a sample relative to the boundary. The classifier trained on filtered and modelled truck signals of the highest mileage truck correctly distinguished all 4 DG and EULtrucks. A model-informed strategy with accurate NH3 storage and TPNOx model estimates gave more accurate results than a purely data-based strategy. The best case result gave one in eight falsely classified day-files when storage was used whereas 4 in eight falsely classified day-files when storage was not used. It is crucial to identify and use a data subset that shows dominant aging signatures. Hence, data filtering strategies play a crucial role in determining the dependability and accuracy of the results.

Degree

M.Sc.

Advisors

Chen, Purdue University.

Subject Area

Aging|Artificial intelligence|Computer science|Mathematics

COinS