近红外光谱结合蒙特卡洛交互验证奇异样本筛选的橄榄油掺伪定性定量分析

Qualitative and quantitative analysis of olive oil adulteration by laser near infrared spectroscopy based on Monte Carlo cross validation

采用基于蒙特卡洛交互验证(MCCV)奇异样本筛选的近红外光谱技术结合支持向量机(SVM)对橄榄油掺伪进行定性和定量分析。应用近红外光谱仪采集将大豆油、菜籽油、玉米油、葵花籽油掺入橄榄油中的188个掺伪样本光谱图。采用蒙特卡洛交互验证(MCCV)方法剔除橄榄油掺伪样本光谱数据中的奇异样本,剔除3个奇异样本。利用多元散射校正(MSC)、去趋势技术(DT)、标准正态变量变换和去趋势技术联用算法(SNV-DT)分别对奇异样本筛选前后的掺伪样本光谱数据进行预处理,选择网格搜索算法(GS)对模型参数组合(C,g)进行寻优,确定最优参数组合。应用支持向量机分类(SVC)方法建立掺伪油的品种定性分类校正模型;采用竞争性自适应重加权算法(CARS)选择奇异样本筛选前后的掺伪样本光谱数据的特征变量,应用支持向量机回归(SVR)建立掺伪油含量快速预测的定量校正模型。试验表明,采用MCCV方法剔除奇异样本后,建立的掺伪油品种鉴别模型的预测准确率达到100%,而建立的GS-SVR模型能够快速预测橄榄油掺伪量,特别是建立SNV_DT-CARS-SVR模型的校正集和预测集相关系数R分别达到99.80%、99.13%,均方误差(MSE)分别为0.0142、0.0535,综合性能最好。结果表明,采用激光近红外光谱分析技术可以实现橄榄油掺伪的定性-定量分析。

Qualitative and quantitative analyses of olive oil adulteration were constructed in this study, by combining near infrared（NIR） spectroscopy based on Monte Carlo cross validation（MCCV） with support vector machines. The spectral data of 188 oils samples was collected by NIR spectrometer, which was made by selecting olive oil soybean oil as base oil, and rapeseed oil, corn oil, sunflower seed oil as adulterated oil. Spectral data that 3 outliers had been eliminated by MCCV method was preprocessed by multiplicative scatter correction（MSC）, de-trending（DT）, standard normal variate transformation deTrending（SNV-DT）, then through the grid search algorithm（GS） combination of model parameter（C, g） was optimized, to determine the optimal parameter combination. Qualitative classification correction model of adulterated oil was set up by SVC method, while quantitative calibration model was done by SVR model after characteristic wavelengths were extracted by competitive adaptive reweighted sampling（CARS）. Experiment showed that after eliminating 3 outliers by MCCV method, model prediction accuracy of qualitative classification could reach 100%, and GS-SVR model could also predict olive oil adulteration quantity quickly, especially for SNV DT-CARS-SVR model, correlation of calibration set and prediction set reached 99.80% and 99.13% respectively, mean square error was 0.0142, 0.0535, whose comprehensive performance was the best. The results showed that the laser near infrared spectroscopy technology can achieve quantitative and qualitative analysis of olive oil adulteration.