Improving NO x emission prediction in ICEVs by recovering aftertreatment-induced time-delayed OBD correlations via sequence-to-sequence approach

Abstract

Precise nitrogen oxide (NO _x ) emission prediction models with highly accessible data are the basis for the development of advanced exhaust gas after-treatment (AFT) systems for engine-powered vehicles. This study proposes a sequence-to-sequence deep learning architecture for predicting real driving emissions (RDE) of NO _x in ICEVs equipped with a state-of-the-art AFT configuration consisting of a lean NO _x trap (LNT) and a selective catalytic reduction (SCR) unit. The proposed model is trained exclusively with onboard diagnostic (OBD) signals, which are universally available in production vehicles and encapsulate comprehensive information on engine operation. To investigate the degradation in prediction accuracy across the AFT system, mutual information analysis is conducted to statistically quantify the loss of correlation between engine-related OBD signals and tailpipe NO _x emissions. The results reveal that the de-NO _x process exhibits highly nonlinear and randomly delayed behavior, resulting in increased temporal uncertainty and reduced model observability. To address these limitations, sequence-to-sequence models incorporating encoder–decoder and attention mechanisms are implemented and further tuned using insights derived from the chemical kinetics of NO _x reduction reactions. The developed model demonstrated superior predictive accuracy and extrapolation capability relative to conventional data-driven baselines under transient and untrained operating conditions. The exclusive use of standardized OBD inputs confirms the feasibility of applying the proposed framework to practical engineering applications, including emissions monitoring, diagnostics, and calibration in on-road situations.