近红外光谱结合化学计量学对五元调和油中单组分油的定量分析

有关调和油快速准确定量检测的研究对于调和油质量控制具有重要意义。以往对调和油定量分析的研究大多集中于二元、三元和四元调和油，对更高元数调和油的研究很少，难以满足调和油检测需求。该研究的目的是探讨近红外光谱结合化学计量学对五元调和油中各单组分油进行定量分析的可行性。由玉米油、大豆油、稻米油、葵花油和芝麻油配制成51个五元调和油样品，并采集各样品12 000～4 000 cm-1范围内的近红外透射光谱。首先，采用光谱-理化值共生距离（SPXY）算法将调和油样品划分为38个校正集和13个预测集样品。其次，考察了主成分回归（PCR）、偏最小二乘（PLS）、支持向量回归（SVR）、人工神经网络（ANN）、极限学习机（ELM）等五种多元校正方法对五元调和油各组分定量分析的建模效果。然后，在最佳建模方法的基础上比较了SG平滑、标准正态变量（SNV）、多元散射校正（MSC）、一阶导数（1st Der）、二阶导数（2nd Der）和连续小波变换（CWT）六种光谱预处理方法，并讨论了预处理方法有效地原因。最后，在最佳预处理方法的基础上进一步利用竞争自适应重加权采样（CARS）和蒙特卡罗无信息变量消除法（MCUVE）筛选与预测组分相关的变量。结果显示，在五种建模方法中，PLS是最佳的建模方法，对玉米油、大豆油、稻米油、葵花油和芝麻油五种组分的预测均方根误差（RMSEP）分别为5.564 4，5.559 2，3.592 6，7.421 8和4.193 0。经过光谱预处理-变量选择，再建立PLS模型，对五种组分的RMSEP分别降低至1.955 3，0.562 4，1.145 0，1.619 0和1.067 1，预测相关系数（R_p）均高于0.98，表明采用合适的光谱预处理和变量选择方法，可以明显提高五元调和油中各单组分油定量分析的预测准确度。该研究为多组分调和油的快速无损定量检测提供了一种参考。

Quantitative Analysis of Single Component Oils in Quinary Blend Oil by Near-Infrared Spectroscopy Combined With Chemometrics

The rapid and accurate quantitative analysis of blend oil is of great importance for the quality control of blend oil. However, most previous studies on the quantitative analysis of blend oil have focused on binary, ternary and quaternary blends, and few studies have been conducted on more multi-component blend oil, which is difficult to meet the needs of blend oil detection. This study explores the feasibility of near infrared spectroscopy combined with chemometrics for the quantitative analysis of the singlecomponentoil in quinary blend oil. 51 quinary blend oil samples were formulated from corn oil, soybean oil, rice oil, sunflower oil and sesame oil, and their NIR spectra were measured in a transmittance mode in the range of 12 000～4 000 cm-1. Firstly, the sample set partitioning based on joint x-y distances (SPXY) algorithm was used to divide the sample into 38 calibration and 13 prediction set samples. Secondly, the modeling effect of five multivariate calibration methods, including principal component regression (PCR), partial least squares (PLS), support vector regression (SVR), artificial neural network (ANN), and extreme learning machine (ELM), were examined for the quantitative analysis of each component in quinary blend oil. Then six spectral preprocessing methods including Savitzky Golag smoothing(SG smoothing), standard normal variate (SNV), multiplicative scatter correction (MSC), first derivative (1st Der), second derivative (2nd Der), and continuous wavelet transform (CWT) were compared based on the best modeling method and the reasons for the effectiveness of the preprocessing methods were discussed. Finally, based on the optimal preprocessing method, the competitive adaptive reweighted sampling (CARS) and Monte Carlo uninformative variable elimination (MCUVE) algorithms were further used to screen the variables associated with the predicted components. The results showed that PLS was the optimal modeling method among the five modeling methods, with root mean square error of the prediction set (RMSEP) of 5.564 4, 5.559 2, 3.592 6, 7.421 8, and 4.193 0 for the five components of corn oil, soybean oil, rice oil, sunflower oil, and sesame oil, respectively. After preprocessing-variable selection and then PLS modeling, the RMSEP for the five components were 1.955 3, 0.562 4, 1.145 0, 1.619 0 and 1.067 1, respectively and the correlation coefficients of prediction set (R_p) were all higher than 0.98, indicating that with appropriate spectral preprocessing, variable selection and modeling methods, the accuracy of quantitative analysis of each component in quinary blend oil was greatly improved. This research provided a reference for rapid and non-destructive quantitative detection of multi-component blend oil.