油脂中的反式脂肪酸(TFA)有害人们的身体健康, 有必要对其含量进行监测。 共收集各类食用植物油样本79个, 涉及9个品种和27个品牌, 分配到校正集和预测集的样本数分别为53个和26个。 采用QE65000拉曼光谱仪采集79个样本的拉曼光谱, 利用自适应迭代惩罚最小二乘法去除样本拉曼光谱的荧光背景; 在此基础上, 采用多种归一化方法对样本拉曼光谱进行处理, 并对拉曼光谱的建模波数范围进行初选; 再利用竞争性自适应重加权采样(CARS)方法筛选与食用植物油TFA含量相关的光谱变量, 并应用偏最小二乘(PLS)回归将食用植物油TFA的特征变量光谱强度与气相色谱测定的TFA真实含量进行关联, 建立食用植物油中TFA含量的定量预测模型。 研究结果表明, 多种归一化方法中, 有4种归一化方法均能提高PLS定量预测模型的性能, 其中Area normalization方法的效果最优; 经建模波数范围初选, 波数范围由686~2 301 cm-1缩减为737~1 787 cm-1, 确定较优的建模波数范围为737~1 787 cm-1; 经CARS方法筛选, 共有31个光谱变量被选择, 其选择的光谱变量主要分布在1 265, 1 303, 1 442及1 658 cm-1拉曼振动峰附近, 且974 cm-1拉曼振动峰两侧均有光谱变量被选择; 此外, CARS方法的PLS建模结果优于常用的无信息变量消除及连续投影算法。 由此可知, 激光拉曼光谱技术结合化学计量学方法检测食用植物油中的TFA含量是可行的。 归一化方法、 建模波数范围初选及竞争性自适应重加权采样(CARS)方法能有效提高TFA定量预测模型的预测精度和稳定性, 优化后的TFA定量预测模型的校正集及预测集的相关系数和均方根误差分别为0.949, 0.953和0.188%, 0.191%。 与未优化的预测模型相比, 预测均方根误差由0.361%下降为0.191%, 下降幅度为47.1%; 建模所用的变量数由683个下降为31个, 仅占原变量数的4.54%。

Trans fatty acids (TFA) in oils and fats are harmful to people’s health, so it is necessary to monitor their content. In this research, 79 samples of edible vegetable oils were collected, involving 9 varieties and 27 brands. The number of samples that were allocated to calibration and prediction sets was 53 and 26, respectively. Raman spectra of 79 edible vegetable oil samples were collected by a QE65000 Raman spectrometer, and adaptive iteratively reweighted penalized least squares was used to remove fluorescence background of Raman spectra. Then, various normalization methods were used to process Raman spectra, and preliminary selection of modeling wavenumber range of Raman spectra was carried out. After that, competitive adaptive reweighted sampling (CARS) method was used to select TFA-related variables, and partial least squares regression was used to correlate the spectral intensity of TFA characteristic variables with the real content determined by gas chromatography to establish quantitative prediction model of TFA content in edible vegetable oils. The results indicate that among various normalization methods, four normalization methods can improve the performance of PLS quantitative prediction model, and area normalization method has the best effect. After primary selection of wavenumber range, the range of wavenumber is reduced from 686 to 2 301 cm-1 to 737 to 1 787 cm-1, and the optimum range of wavenumber is determined to be 737 to 1 787 cm-1. Thirty-one spectral variables are selected by CARS method. The selected spectral variables are mainly distributed near the Raman vibration peaks of 1 265, 1 303, 1 442 and 1 658 cm-1, and the variables in the both sides of the Raman vibration peaks of 974 cm-1 are also selected. In addition, the PLS modeling results of CARS method were better than those of the commonly used methods such as uninformative variable elimination and successive projections algorithm. Therefore, it is feasible to detect TFA content in edible vegetable oil by laser Raman spectroscopy combined with chemometrics. Normalization method, wavenumber range selection and CARS method can effectively improve the prediction accuracy and stability of TFA quantitative prediction model. The correlation coefficients and root mean square errors of optimized TFA quantitative prediction model in calibration and prediction sets are 0.949, 0.953 and 0.188%, 0.191%, respectively. Compared with the unoptimized prediction model, the root mean square error of prediction decreases from 0.361% to 0.191%, with a decrease of 47.1%. The number of variables used in modeling decreases from 683 to 31, accounting for only 4.54% of the original variables.