基酒FT-NIR光谱预处理与特征波筛选方法的比较  

Comparison of FT-NIR Spectral Pretreatment and Characteristic Band Screening for Baijiu-based Liquor

作　　者：朱雪梅 庹先国 张贵宇 翟双 罗林 罗琪 ZHU Xuemei;TUO Xianguo;ZHANG Guiyu;ZHAI Shuang;LUO Lin;LUO Qi(School of Automation&Information Engineering,Artificial Intelligence Key Laboratory of Sichuan Province,Sichuan University of Science&Engineering,Yibin 644000,China;School of Information Engineering,Southwest University of Science and Technology,Mianyang 621010,China)

机构地区：[1]四川轻化工大学自动化与信息工程学院,人工智能四川省重点实验室,四川宜宾644000 [2]西南科技大学信息工程学院,四川绵阳621010

出　　处：《现代食品科技》2023年第1期196-204,共9页Modern Food Science and Technology

基　　金：四川省科技计划项目(2022YFS0554);四川省重大科技专项项目(2018GZDZX0045);四川省科技成果转移转化示范项目(2020ZHCG0040)。

摘　　要：为解决白酒基酒分类的问题,降低基酒的分类误差,减少基酒对摘酒师傅身体的危害,本实验选取18种预处理以及3种特征波筛选方法来较少光谱中的无关干扰信息,降低建模数据复杂度。基酒的傅里叶近红外光谱(Fourier Transform Near Infrared Spectroscopy,FT-NIR)经过光谱理化值共生距离法(SPXY)划分数据集、预处理、马氏距离(MD)异常剔除、特征波筛选、支持向量机回归(SVR)预测来完成最终的分类。研究发现:多元散射校正(Multiplicative Scatter Correction,MSC)后的训练集预测集分类准确率可以达到100%,主成分分析(Principal Component Analysis,PCA)与特定算法结合才能实现准确分类,因此要注意与其他算法的组合,无信息变量消除法(Uninformative Variables Elimination,UVE)和竞争性自适应重加权算法(Competitive Adaptive Reweighted Sampling,CARS)都能实现高效的特征波选择,预测集的平均准确率接近90%。实验证明,经过处理后的光谱数据最多占原数据的47.57%,基酒近红外谱图经过预处理与特征波筛选后可以降低后期回归模型处理数据的复杂程度,提高模型的精确度。To classify baijiu base wine,reduce the classification error of baijiu base wine,and reduce the harm of base wine to the body of Baijiu Based Liquor pickers,18 pretreatment methods and three characteristic wave screening methods were selected to reduce irrelevant interference information in the spectrum and complexity of the modeling data.The Fourier-transform near infrared spectra of baijiu base wine were divided into datasets using SPXY and preprocessed,and then subjected to Mahalanobis distance anomaly elimination,eigenwave screening,and support vector machine regression prediction.After multiplicative scatter correction,the classification accuracy of training set prediction was 100%.Principal component analysis can be combined with specific algorithms to achieve accurate classification;studies are needed to combine this analysis with other algorithms.Uninformative variables elimination and competitive adaptive reweighted sampling can achieve efficient feature wavelength extraction,with an average accuracy of prediction of close to 90%.The experimental results showed that the processed spectral data accounted for up to 47.57%of the original data,the complexity of the regression model was reduced,and the accuracy of the model was improved after pretreatment and characteristic wavelength selection of the near-infrared spectrum of the base wine.

关 键 词：近红外 基酒分级 多元散射校正 无信息变量消除法 竞争性自适应重加权算法 

分 类 号：TS262.3[轻工技术与工程—发酵工程] O657.33[轻工技术与工程—食品科学与工程]