拉曼光谱结合模式识别方法用于大豆原油掺伪的快速判别

大豆原油是我国的战略储备物资，然而目前储油市场上频繁出现大豆原油掺混的现象严重影响了食用油储备安全。基于此，通过大豆原油与部分植物精炼油拉曼谱图的特征差异，并结合主成分分析-支持向量机(PCA-SVM)模式识别建立了大豆原油是否掺伪的快速判别方法。以28个大豆原油、46个精炼油、110个掺伪油的拉曼谱图为模型样本;选择位于780～1 800 cm-1波段的谱图，预处理方法同时采用Y轴强度校正、基线校正和谱图归一化法;在此基础上应用PCA法提取特征变量，即以贡献率最高前7个主成分为变量进行SVM分析。SVM校正模型的建立是以随机选取的20个大豆原油和75个掺伪油样组成校正集，以8个大豆原油和35个掺伪油样组成验证集，分别运用并比较四种核函数算法建立的大豆原油SVM分类模型，并采用网格搜索法(grid-search)优化模型的参数，以四种模型的分类性能作为评判标准。结果表明：应用线性核函数算法构建的SVM分类模型可以很好地完成掺伪大豆原油的判别，校正集识别准确率达到100%，预测结果的误判率为0，判别下限为2.5%。结果表明应用拉曼光谱结合化学计量学能够用于大豆原油掺伪的快速鉴别。拉曼光谱简便、快速、无损、几乎没有试剂消耗，适合现场检测，从而为大豆原油的掺伪分析提供了一种新的备选方法。

Raman Spectroscopy Combined with Pattern Recognition Methods for Rapid Identification of Crude Soybean Oil Adulteration

In the present paper, a non-destructive, simple and rapid analytical method was proposed based on Raman spectroscopy (Raman) combined with principal component analysis (PCA) and support vector machine (SVM) as pattern recognition methods for adulteration of crude soybean oil (CSO). Based on fingerprint characteristics of Raman, the spectra of 28 CSOs, 46 refined edible oils (REOs) and 110 adulterated oil samples were analyzed and used for discrimination model establishment. The preprocessing methods include choosing spectral band of 780~1 800 cm-1, Y-axis intensity correction, baseline correction and normalization in succession. After those series of spectral pretreatment, PCA was usually employed for extracting characteristic variables of all Raman spectral data and 7 principal components which were the highest contributions of all data were used as variables for SVM model. The SVM discrimination model was established by randomly picking 20 CSOs and 95 adulterated oils as calibration set, and 8 CSOs and 35 adulterated oils as validation set. There were 4 kinds of kernel function algorithm (linear, polynomial, RBF, sigmoid) respectively used for establishing SVM models and grid-search for optimization of parameters of all the SVM models. The classification results of 4 models were compared by their discrimination performances and the optimal SVM model was based on linear kernel classification algorithm with 100% accuracy rate of calibration set recognition, a zero misjudgment rate and the lowest detection limit of 2.5%. The above results showed that Raman combined PCA-SVM could discriminate CSO adulteration with refined edible oils. Since Raman spectroscopy is simple, rapid, non-destructive, environment friendly, and suitable for field testing, it will provide an alternative method for edible oil adulteration analysis.