基于机器学习的柴油机纳米级微粒预测模型

预测柴油机燃烧产生的纳米级微粒是减少空气污染的有效方法之一,为车辆颗粒物(particulate matter, PM)的排放监管与控制提供支持,协助标定工程师制定严格的排放法规。采用气缸压力传感器测量柴油机在不同行驶工况下的气缸压力,利用主成分分析(principal component analysis, PCA)方法提取前4、7、10主成分作为神经网络的训练输入,粒径为7～990 nm的颗粒物浓度作为模型的输出,分析不同工况下气缸压力主成分贡献率对纳米颗粒的预测效果。结果表明:利用较少的主成分即可代表不同工况下的缸压燃烧特性;当主成分贡献率达到91.57%时,粒径为7～990 nm的颗粒物浓度试验数据与模型预测的平均绝对误差为90.74 cm~3,均方根误差为1.612×10~4 cm~3,回归系数R~2达到0.95,预测精度较高。因此,利用气缸压力预测柴油机PM的排放是一种可行方案。

Diesel Nano-scale Particle Prediction Model Based on Machine Learning

Predicting the nanoscale particles produced by diesel engine combustion is one of the effective methods to reduce air pollution. It supports the monitoring and control of vehicle particulate emis-sions and assists calibration engineers in setting strict emission regulations. The cylinder pressure sensor was used to measure the cylinder pressure of diesel engine under different driving conditions. The first 4,7,10 principal components were extracted by principal component analysis method as the training input of neural networks, and the numerical concentration of 7~990nm particle size range was used as the test output. The predictive effect of the principal component contribution rate of cyl-inder pressure on nanoparticles under different working conditions was analyzed. The results show that the variation characteristics of cylinder pressure under different working conditions can be fully expressed by the principal component coefficient (PCA). At the same time, when the principal com-ponent reaches a certain proportion, the average absolute error between the measured and predictive output of particle number concentration with particle size 7~990nm is 90.74/cm3, the root mean square error is 16121.10/cm3, and the regression coefficient R2 is 0.95. Therefore, it is a feasible scheme to predict the emission of PM by the cylinder pressure generated by diesel engine combus-tion.