MIV波长优选改善VIS/NIR光谱TVB-N模型性能研究  

Improvements of VIS-NIR Spectroscopy Model in the Prediction of TVB-N Using MIV Wavelength Selection

作　　者：陈亦凡 李芸婧 彭苗苗 杨春勇[1] 侯金[1] 陈少平[1] CHEN Yi-fan;LI Yun-jing;PENG Miao-miao;YANG Chun-yong;HOU Jin;CHEN Shao-ping(Hubei Key Laboratory of Intelligent Wireless Communications,College of Electronics and Information Engineering,South-Central University for Nationalities,Wuhan 430074,China)

机构地区：[1]智能无线通信湖北省重点实验室,中南民族大学电子信息工程学院,湖北武汉430074

出　　处：《光谱学与光谱分析》2020年第5期1413-1419,共7页Spectroscopy and Spectral Analysis

基　　金：湖北省自然科学基金科技支撑计划项目(2015BCE048);国家科技支撑计划课题(2015BAD29B01);湖北省自然科学基金重点项目(2014CFA051);中央高校基本科研业务费专项资金自科培育项目(CZP17026)资助。

摘　　要：挥发性盐基氮(TVB-N)是衡量肉品新鲜的重要理化指标,利用可见/近红外(VIS/NIR)光谱对TVB-N含量进行定量检测具有重要意义。预测模型是VIS/NIR光谱检测TVB-N含量性能的关键要素,使其兼顾准确性与稳健性可有效改善TVB-N的定量分析结果。以猪肉为例,采集51组不同新鲜度样本的VIS/NIR光谱数据,去除低信噪比区间200~450和900~1000 nm,选取有效波段450~900 nm的光谱数据用于建模。随后利用主成分分析(PCA)对光谱信息降维,构建一个反向传播神经网络(BPNN)模型。在此基础上,提出用平均影响值(MIV)方法从有效波段中优选与肉质TVB-N含量强相关的特征波长,最终基于221个优选波长,构建一个MIV-PCA-BPNN预测模型。实验表明,初步构建的PCA-BPNN非线性预测模型,校正相关系数(R_C)和校正均方根误差(RMSEC)分别为0.96和1.47 mg/100 g,预测相关系数(R_P)和预测均方根误差(RMSEP)分别为0.93和1.74 mg/100 g,模型稳健性指标为1.18,优于经典的线性预测模型主成分分析回归和偏最小二乘回归,证明TVB-N具有较强的非线性效应。最终构建的MIV-PCA-BPNN预测模型的R_C和RMSEC分别为0.98和1.21 mg/100 g,R_P和RMSEP分别为0.96和1.12 mg/100 g,模型稳健性指标为1.08,在所构建的预测模型中,RMSEC和RMSEP最小,RC和RP最大,模型的准确性和稳健性最佳。另外,MIV方法筛选出的特征波长集中在7个波峰附近,皆分布于肉品中化学成分的吸收区内,且与TVB-N中的含氢基团的特征吸收峰表现出高度一致性,为利用MIV方法筛选波长变量提供了理论依据。研究结果显示,MIV波长优选可有效改善预测模型的性能,为利用神经网络剔除无关波长变量提供了新思路,所构建的MIV-PCA-BPNN预测模型满足了肉质中TVB-N定量分析的需求。Volatile Basic Nitrogen(TVB-N)is an important physicochemical property for the detection of meat freshness.Using visible/near-infrared(VIS/NIR)spectroscopy to analyze TVB-N content is of great importance quantitatively-.The prediction model is the key factor for detection TVB-N content in visible or near infrared spectroscopy.Thus,an accurate and robust prediction model can improve the quantitative analysis results of TVB-N.Firstly,we collected 51 representative pork samples with different freshness,and determine the effective band from 450 to 900 nm after removing low signal-to-noise ratio band from 200 to 450 nm and from 900 to 1000 nm.Then we use principal component analysis(PCA)to reduce spectral data in order to construct a back propagation neural network(BPNN)model.On this basis,we use the mean impact value(MIV)method to select characteristic wavelengths which strongly related to the content of Total Volatile Basic Nitrogen(TVB-N)in edible meat,and finally construct a MIV-PCA-BPNN prediction model based on 221 selected wavelengths.Experimental results show that the related coefficient of calibration(R_C),the related coefficient of prediction(R_P),the root means square error of calibration(RMSEC),the root mean square error of prediction(RMSEP)and the robustness index of the PCA-BPNN model are 0.96,0.93,1.47 mg/100 g,1.74 mg/100 g and 1.18,respectively.The PCA-BPNN nonlinear prediction model is better than the classical linear prediction model principal component regression and partial least squares regression prediction model,which proves that TVB-N has strong nonlinear effects.The R_C,R_P,RMSEC,RMSEP and the robustness index of the MIV-PCA-BPNN model are 0.98,0.96,1.12 mg/100 g,1.21 mg/100 g and 1.08,respectively,it is RMSEC and RMSEP are the smallest,while RC,RP are the largest.Therefore,MIV-PCA-BPNN is the most accurate and robust model in all constructed prediction model.In addition,the characteristic wavelengths selected by the MIV method are concentrated near 7 peaks,which are distributed in the absorp

关 键 词：VIS/NIR光谱检测 反向传播神经网络 波长优选 挥发性盐基氮 

分 类 号：O433.4[机械工程—光学工程]