中红外光谱法结合支持向量机快速鉴别蜂蜜品种  

Mid-Infrared Spectroscopy Analysis Combined with Support Vector Machine for Rapid Discrimination of Botanical Origin of Honey

作　　者：徐天扬[1,2,3] 杨娟 孙晓荣[4,5] 刘翠玲[4,5] 李熠 周金慧[1,2,3] 陈兰珍 Xu Tianyang;Yang Juan;Sun Xiaorong;Liu Cuiling;Li Yi;Zhou Jinhui;Chen Lanzhen(Institute of Apicultural Research, Chinese Academy of Agricultural Sciences, Beijing 100093, China;Key Laboratory of Bee Products for Quality and Safety Control, Ministry of Agriculture, Beijing 100093, China;Laboratory of Risk Assessment for Quality and Safety of Bee Products, Ministry of Agriculture, Beijing 100093, China;School of Computer and Information Engineer, Beijing Technology and Business University, Beijing 100048, China;Beijing Key Laboratory of Large Data Technology for Food Safety, Beijing 100048, China)

机构地区：[1]中国农业科学院蜜蜂研究所,北京100093 [2]农业部蜂产品质量安全控制重点实验室(北京),北京100093 [3]农业部蜂产品质量安全风险评估实验室,北京100093 [4]北京工商大学计算机与信息工程学院,北京100048 [5]食品安全大数据技术北京市重点实验室,北京100048

出　　处：《激光与光电子学进展》2018年第6期425-433,共9页Laser & Optoelectronics Progress

基　　金：国家自然科学基金面上项目(331772070);中国农业科学院创新工程项目(CAAS-ASTIP-2017-IAR);国家特色农产品风险评估专项(GJFP2017010);国家蜂产业技术体系(CARS-45-KXJ10)

摘　　要：为快速鉴别5种蜂蜜(椴树蜜、荆条蜜、油菜蜜、洋槐蜜、荔枝蜜)的品种,首次提出了基于主成分分析(PCA)方法结合线性支持向量机(SVM)或最小二乘支持向量机(LSSVM)的中红外光谱法鉴别蜂蜜品种的新方法。用傅里叶变换中红外光谱仪测定5种蜂蜜样本的中红外光谱,并进行归一化预处理,然后用主成分分析降维方法分别提取经预处理后的光谱数据中的5维、10维、15维、20维特征数据,最后设计了线性SVM和基于网格搜索优化算法的径向基函数(RBF)的LSSVM分类器模型。利用不同分类器模型,识别未知蜂蜜样本光谱数据降维到不同维数的特征数据,并进行实验验证。结果表明:应用主成分分析降维方法降维到20维的特征数据在SVM和LSSVM分类器上的平均识别率均高于97%,最高识别率均可达到100%,且稳定性很好;利用较低维数数据进行分类时,LSSVM分类器比SVM的识别精度更高,稳定性更好。研究证明将中红外光谱与线性SVM或LSSVM结合用于快速鉴别蜂蜜品种是可行的。To achieve the fast discrimination of five varieties of honeys, namely linden honey, vitex honey, rape honey, acacia honey and litchi honey, we propose a new method in this article by using the mid-infrared spectra based on principle component analysis （PCA） combined with linear support vector machine （SVM） or least squares support vector machine （LSSVM）. The mid-infrared spectra of five varieties of honey samples are determined by Fourier transform infrared spectroscopy and normalized. Then the 5-dimensional, 10-dimensional, 15-dimensional, and 20-dimensional feature data will be extracted from spectra with the use of dimension reduction method of PCA after normalization. Finally, the two classifier models, linear SVM and LSSVM with radial basis function （RBF） based on the grid search optimization, are designed. Using different classifier model, we identify the different dimensional feature data extracted from spectra data of unknown honey samples. Then the results of different dimension feature data and different support vector machines are validated. Experimental results show that for the 20-dimensional feature data obtained by the dimension reduction method of PCA, an average recognition rate of higher than 97% on SVM and LSSVM classifiers is achieved, the highest recognition rate can reach 100%, and classifier stability is very good. LSSVM classifier has higher recognition accuracy and better stability than linear SVM classifier in classification with lower dimension data. Hence, it proves the feasibility of rapid identification of five varieties of honeys with mid-infrared spectra combined with linear SVM or LSSVM.

关 键 词：光谱学 中红外光谱 主成分分析 支持向量机 最小二乘支持向量机 径向基函数 

分 类 号：O657.33[理学—分析化学]