Modeling real-world driving emissions of a plug-in hybrid electric vehicle by multi-domain adversarial training

Real-world driving emission (RDE) tests cannot cover all the diverse driving scenarios, so the real-world driving emission models trained with the RDE test datasets tend to demonstrate localized overfitting and inferior generalization performance. By transfer learning concepts, this study identified the emission-related feature distribution bias in different rounds of RDE tests as domain distinctiveness, and developed a multi-domain adversarial neural network (Multi-DANN) architecture to model real-world driving emissions. Using multi-domain adversarial training, the Multi-DANN model leveraged the distinct emission-related feature distributions across various source and target domains, extracting domain-invariant features most relevant to real-world driving emissions. This approach improved the model's prediction ability in the target domain. Six rounds of RDE test datasets of a plug-in hybrid vehicle were used for the multi-domain adversarial training and testing. Domain rotation cross-testing reveals that the Multi-DANN model achieves better prediction accuracy and generalization performance in target domains than models employing traditional long short-term memory and domain adversarial neural network architectures. Additionally, the Multi-DANN model excels at capturing outlier features through adequate feature alignment, thereby demonstrating considerable advantages in predicting abnormal pulse PN and CO emissions caused by engine start-stop events in the plug-in hybrid electric vehicle.