Transfer learning for transient NOx, PN and THC emission prediction of non-road diesel engines based on NRTC experiments.

Abstract

This study introduces a novel task transfer learning framework for predicting transient emissions (NOx, PN, and THC) in non-road diesel engines. Our key innovation lies in eliminating model re-optimization through a fixed-architecture approach where pretrained hyperparameters are preserved and only task-specific layers are fine-tuned. Validated on NRTC data across all emission transfer scenarios, the method achieves near-identical accuracy to pretrained models (R² difference ≤0.0044), peak R² values of 98.87% (NOx), 99.54% (PN), and 99.52% (THC) and computational cost reduction by 72% versus conventional methods. The framework surpasses operational vehicle sensor accuracy and matches laboratory-grade equipment precision. Analysis confirms the efficacy of transfer learning for emission prediction and establishes an efficient pre-trained model organization paradigm.