多维标度和随机森林应用于汽油近红外定量分析

随机森林(RF)回归应用于汽油辛烷值的近红外定量模型的波长优选。提出的双中心指标多维标度(DC-MDS)方法能够有效地找到定标和预测样品集的合理划分。RF回归建模的过程中选择采用较大的决策树数量(nTree=500),避免建模过程发生拟合,进一步调试并选择最优的分裂变量数(mtry=130);最后在最优参数的RF建模过程中提取具有最大贡献的30个信息波长,为汽油辛烷值的测定建立离散波长的近红外定量分析模型;其预测决定系数为0.971,预测均方根偏差为0.219%。结果表明,RF回归具有应用于汽油辛烷值近红外定量测定的潜力。

Near-Infrared Spectroscopic Determination for Gasoline by using Multi-Dimensional Scaling and Random Forest

Random forest (RF) regression applied to the wavelength selection for the near-infrared (NIR) quantitative analysis of gasoline octane. A sample division method called duel-center multi-dimensional scaling (DC-MDS) was proposed to search a reasonable division of calibration and prediction sample sets. We utilized a large number of decision trees (nTree=500) for RF regression, in order to avoid the occurrence of over-fitting; and further had the number of split valuables tunable, searching for the optimal value (m_try=130). And we also obtained 30 informative wavelengths in the process of RF modeling, which were expected for the establishment of a NIR model for gasoline octane, based on seldom discrete wavelengths. This optimal model output the predictive determination coefficient of 0.971and a root mean square root of 0.219 (%). The results showed that, RF regression has the potential to be applied to the NIR analysis of gasoline octane.