拉曼光谱和MLS-SVR的食用油脂肪酸含量预测研究

为实现食用植物油中饱和脂肪酸、油酸、亚油酸含量的快速预测，对一批纯食用油以及不同比例两两混合油共91个样品进行了拉曼光谱检测，在800～2 000 cm-1范围内，通过基于寻峰算法的自动确定支点的基线拟合方法，对获得的光谱数据进行预处理，提取八个特征峰作为拉曼光谱的特征值。以这些特征值为输入，以样品油中实际饱和脂肪酸、油酸、亚油酸含量为输出，运用偏最小二乘回归(PLS)和多输出最小二乘支持向量回归机(MLS-SVR)方法，分别建立了可以同时预测三种脂肪酸含量的数学模型，结果表明MLS-SVR方法具有较好的效果。将MLS-SVR模型的预测结果与气相色谱法结果相比较，可得到三种脂肪酸的预测均方根误差分别为0.496 7%，0.840 0%和1.019 9%，相关系数分别为0.813 3，0.999 2和0.998 1;对未知样品三种脂肪酸的预测均方根误差不超过5%。表明，拉曼光谱和MLS-SVR相结合的食用油脂肪酸含量预测方法，具有快速、简便、无损、准确等优点，为食用油脂肪酸含量分析提供了一种可行的方法。

Research on Prediction Method of Fatty Acid Content in Edible Oil Based on Raman Spectroscopy and Multi-Output Least Squares Support Vector Regression Machine

For the purpose of the rapid prediction of saturated fatty acid, oleic acid, linoleic acid content in edible vegetable oil, the Raman spectra of a batch of edible vegetable oils and their one-one mixtures with different ratios were measured in the range of 800~2 000 cm-1, 91 samples were measured totally in this research, the obtained Raman spectra data were preprocessed by a new method proposed in this paper called auto-set fulcrums baseline fitting method based on peak-seeking algorithm, and 8 characteristic peak values (872 cm-1ν(C—C)]，972 cm-1δ(CC)trans]，1 082 cm-1ν(C—C)]，1 267 cm-1δ(C—H)cis]，1 303 cm-1δ(CH₂)twisting]，1 442 cm-1δ(CH₂) scissoring]，1 658 cm-1ν(CC)cis]，1 748 cm-1ν(CO)]) were extracted to be the eigenvalues for the whole spectra, among the 8 peaks there are three peaks(972, 1 267, 1 658 cm-1) that play an important role in the establishment of mathematical model, they are closely concerned with CC band which distinguishes the three fatty acid types. By using these eigenvalues as inputs, and actual saturated fatty acid, oleic acid, linoleic acid contents of sample oils as outputs, a prediction mathematical model that predicts simultaneously the three fatty acid contents was established using multiple regression analysis: multi-output least squares support vector regression machine(MLS-SVR) and partial least squares(PLS). Results show that the MLS-SVR has better effects. The predicting results are compared with results of gas chromatography(GC), and the obtained root mean square error of prediction(RMSEP) for saturated fatty acid, oleic acid, linoleic acid are 0.496 7%, 0.840 0% and 1.019 9%, and the correlation coefficients (r) are 0.813 3, 0.999 2 and 0.998 1, respectively. When this model is applied in the detection of new unknown oil samples, the prediction error does not exceed 5%. Results show that the Raman spectra analysis technology based on MLS-SVR can be a convenient, fast, non-destructive, and precise new method for oil detection.