Improving the prediction accuracy of aircraft engine non-volatile particulate matter emission indices through deep learning model

Accurate prediction of aircraft engine non-volatile particulate matter (nvPM) emissions during the landing and take-off (LTO) cycle is crucial for assessing its environmental impact on airport air quality. However, existing statistical methods recommended by the civil aviation industry rely on the engine smoke number and other indirect influencing parameters. By utilizing the parameter fitting ability of convolutional neural network (CNN) and the parameter selection ability of support vector machine (SVM) under small-sample datasets, an integrated CNN-SVM model to predict aircraft engine nvPM mass and number emission indices in the LTO cycle is proposed in this study. The prediction accuracy of the CNN-SVM model is compared with that of the statistical model, and the predicted nvPM mass and number emission indices for typical aircraft engines are subsequently used to evaluate their emission performance. Results show that over 90% of estimated nvPM mass and number emission indices fall within the 95% prediction interval for all LTO phases, outperforming relevant statistical method with higher R² (0.95 vs. 0.8 and 0.94 vs. 0.82) and lower RMSE (43.85 mg/kg vs. 82 mg/kg and 8.36*10¹⁴ #/kg vs. 1.6*10¹⁵ #/kg) respectively. Additionally, the evaluation of typical engine nvPM emission performance based on predicted values via CNN-SVM model is consistent with relevant certification results, demonstrating its feasibility for engine emission compliance assessment. Therefore, the proposed CNN-SVM model can not only achieve high-accuracy data to assess the environmental impact of aircraft engine nvPM emissions generated in the LTO cycle, but also potentially provide an effective data-driven approach for equivalent airworthiness evaluation of aircraft engine nvPM emissions in the global aviation industry.