基于反射光谱的油茶籽油掺伪量快速测定 及特征波长特性研究Rapid prediction of oil-tea camellia seed oil adulteration amount based on reflection spectroscopy and characteristic wavelength characteristics

为了探索紫外-可见-近红外反射光谱测定油茶籽油掺伪量的方法，按照不同掺伪比例制备了244个油茶籽油掺伪大豆油、菜籽油、花生油、玉米油的样本，以自主搭建的实验平台采集所制备样本在200~1 100 nm范围内的反射光谱。将原始光谱进行Savitzky-Golay(SG)-连续小波变换（CWT）预处理后，利用Kennard-Stone(K-S)算法以2∶ 1的比例将样本划分成校正集和预测集。采用竞争性自适应重加权算法（CARS）、连续投影算法（SPA）、自主软收缩算法（BOSS）、迭代变量子集优化算法（IVSO）进行特征波长选择，分别建立基于支持向量机（SVM）、极限学习机（ELM）、随机森林（RF）的油茶籽油掺伪量快速预测模型，同时对特征波长的特性进行了研究。结果表明：原始光谱经过 SG-CWT（L5）预处理和 BOSS 特征波长筛选后，建立的基于SVM的油茶籽油掺伪量快速预测模型能够鉴别掺伪量为1%及以上的油茶籽油，该模型在十折交叉验证和网格搜索法下得到最佳惩罚因子（c）和核函数（γ）分别为5.278 0和0.108 8，其预测决定系数（R2P）、预测均方根误差（RMSEP）、预测平均绝对误差（MAEP）分别为0.998 5、0.013 4、0.010 2。特征波长聚集程度和陡度对模型预测结果存在一定影响。综上，建立的基于反射光谱的油茶籽油掺伪量快速预测模型预测误差小，预测效果较好。In order to explore the method of UV-Vis-NIR reflection spectroscopy to identify blended oil-tea camellia seed oil(CAO), 244 samples of CAO adulterated with soybean oil, rapeseed oil, peanut oil and corn oil were prepared according to different adulteration amounts, and the reflectance spectra of the prepared samples in the range of 200-1 100 nm were collected by an experimental platform built independently. After pretreating the raw spectra with SG-continuous wavelet transform (CWT), the samples were divided into correction and prediction sets using the Kennard-Stone (K-S) algorithm in a ratio of 2∶ 1. Competitive adapative reweighting sampling(CARS)algorithm, successive projections algorithm (SPA), bootstrapping soft shrinkage(BOSS) algorithm, and iteratively variable subset optimization (IVSO) algorithm were used for characteristic wavelength selection, and rapid identification models based on support vector machine (SVM), extreme learning machine (ELM), and random forest (RF) were established for CAO adulteration amount, respectively, and the characteristics of characteristic wavelength were studied. The results showed that the SVM model established after the SG-CWT (L5) pretreating and BOSS characteristic wavelength screening could discriminate the amount of adulteration 1% and above, and the model obtained the best penalty factor c ( 5.278 0) and kernel function γ (0.108 8) under the ten-fold cross-validation and grid search method, with R2P, RMSEP and MAEP of 0.998 5, 0.013 4 and 0.010 2, respectively. At the same time, the degree of aggregation and steepness of the characteristic wavelength had some influence on the model prediction results. In conclusion, the established rapid prediction model for the adulteration amount of oil-tea camellia seed oil based on reflection spectroscopy has low error and good prediction effect.