作 者：
王海燕;桂冬冬;沙敏;王彦波;程永波;张正勇
单 位：
南京财经大学管理科学与工程学院;浙江工商大学食品与生物工程学院;浙江工商大学管理工程与电子商务学院;江苏省质量安全工程研究院
关键词：
拉曼光谱;模式识别;化学计量学;牛奶制品;参数优化
摘 要：
本试验以牛初乳奶片、乳酸奶片为对象,采集了多个样品的拉曼光谱并进行结构解析,随后以拉曼光谱为输入,先后论证了多种数据预处理方法的适用性,以及特征提取方法的选择、分类算法的参数优化。结果显示,奶片的拉曼光谱可提供样品丰富的化学结构信息,表征蛋白质、脂肪、糖类等营养物质信息,但由于样品间具有较高的谱图相似性,仅凭裸眼无法实现类别判别。引入支持向量机分类器,系统讨论了包括小波降噪、多元散射校正、求导、归一化多种数据预处理方法,结果发现小波降噪、多元散射校正、一阶求导以及归一化相结合的预处理可有效提高分类器识别率。进一步运用主成分分析特征提取算法,揭示出特征提取可提高分类算法的运行效率、去除冗余信息、节省运行时间、提高分类准确率。随后,比较了网格搜索算法、粒子群优化算法、遗传算法,用以支持向量机分类器参数优化,揭示出网格搜索算法可高效获得最佳惩罚参数(c=11.3731)和核函数参数(γ=0.00097656),可供建立优化的智能判别模型。
译 名：
Application of Raman Spectroscopy Combined with Pattern Recognition Algorithm for Intelligent Identification of Dairy Products and Parameter Optimization
作 者：
WANG Hai-yan;GUI Dong-dong;SHA Min;WANG Yan-bo;CHENG Yong-bo;ZHANG Zheng-yong;School of Management Science and Engineering, Nanjing University of Finance and Economics;School of Management Engineering and Electronic Commerce, Zhejiang Gongshang University;Jiangsu Province Institute of Quality and Safety Engineering;School of Food Science and Biotechnology, Zhejiang Gongshang University;
关键词：
Raman spectroscopy;;Pattern recognition;;Chemometrics;;Dairy products;;Parameter optimization
摘 要：
The aim of this work is to develop a new rapid and intelligent discrimination method for analyzing the dairy products with high similarity based on Raman spectroscopy and pattern recognition classification algorithm. The bovine colostrum milk and lactic acid milk were employed as the research objects in this work, Raman spectra of several samples were collected and structure analyses were made. Then, Raman spectra was used as inputs, the applicability of various methods of data preprocessing, feature extraction method and the parameter optimization of classification algorithm had been demonstrated successively. The results showed that the Raman spectra of milk could provide rich information about the chemical structure of samples, and characterize their nutrients such as protein, fat, carbohydrate. However, we could achieve class discrimination using naked eye only, because there was higher spectra similarity between the samples. Next, support vector machine classifier was introduced in this paper, and pretreatment method including wavelet denoising, multiple scattering correction, derivation and normalization had been fully discussed. The research found that wavelet denoising, multiple scattering correction, and 1 st order derivative and normalized combination pretreatment could effectively improve the recognition rate of classifier. Furthermore, the feature extraction algorithm based on principal component analysis was used, the results displayed that the feature extraction could improve the operation efficiency, remove redundant information, save the running time and improve the classification accuracy. Then, the grid search algorithm, particle swarm optimization algorithm and genetic algorithm were applied to optimize the parameters of support vector machine classifier, respectively. The results revealed that the grid search algorithm could efficiently obtain the optimal penalty parameter c=11.3731 and the kernel function parameter γ=0.00097656. Hence, the intelligent optimization model could been established.
