将蒙牛、伊利、完达山三个品牌的奶粉样品掺入不同量的淀粉构成32份实验样品.在跨度近两个月时间内,用JDSU微型近红外光谱仪,分五天重复5次采集这些样品的中波近红外漫反射光谱.采用仿生模式识别(BPR)算法对样品进行掺假识别定性分析,并研究了分析的可靠性与模型的稳健性.以90%作为评价分析结果(样品掺杂的正确识别率 CAR与正确拒识率 CRR)的阈值:将测试结果高于此阈值的所有样品中掺入淀粉的最低含量分别称为样品掺杂的正确识别限与正确拒识限.结果显示:三个品牌奶粉样品分别各自建模时,若用同一天测定的部分光谱数据建立模型,预测该天剩余光谱,样品掺杂的正确识别限与正确拒识限都可以达到0.1%.对于三种品牌奶粉合并后的纯奶粉及其淀粉掺杂样品混合建模时,若用同一天测定的光谱建模与测试,样品掺杂的正确识别限也可以达到0.1%,正确拒识限则为1%;若用不同时间采集的光谱进行交叉测试,正确识别限与正确拒识限都只有5%;若用四天的光谱数据联合建模,测试第五天的数据,正确识别限可以稳定达到1%,正确拒识限可以达到5%.应用两种算法对奶粉中淀粉含量进行定量分析比较,进一步验证了有关定性分析对样品掺杂正确识别限和正确拒识限的可靠性.

Three China trademarks of milk powder called Mengniu,Yili,Wandashan were taken as testing samples.Each of them mixed varied amount of starch in different gradient,which were consisted of 32 adulterated milk powder samples mixed with starch,was taken as standard samples for constructing predicted model.To those 32 samples,the reflecting spectrum characteristics in middle wave of near infrared spectrum with Near Infrared Spectrum Analyzer (Micro NIR 1700) produced by JDSU Ltd.USA were collected for five repeats in five different days.The time span was nearly two months.Firstly,we build the model used the reflecting spectrum characteristics of those samples with biomimetic pattern recognition (BPR) arithmetic to do the qualitative analysis.The analysis included the reliability of testing result and stability of the model.When we took ninety percent as the evaluation threshold of testing result of CAR(Correct Acceptance Rate)and CRR (Correct Rejection Rate),the lowest starch content of adulterate milk powder in all tested samples which the tested result were bigger than that abovementioned threshold was designated CAR threshold (CAR-T) and CRR threshold (CRR-T).CAR means the correct rate of accepting a sample which is belong to itself,CRR means correct rate of refusing to accept a sample which is not belong to itself.The results were shown that,when we constructed a model based on the near infrared spectrum data from each of three China trademark milk powders,respectively,if we constructed a model with infrared spectrum data tested in a same day,both the CAR-T and CRR-T of adulterate starch content of a sample can reach 0.1% in predicting the remainder infrared spectrum data tested within a same day.The three China trademarks of milk powder had the same result.In addition,when we ignored the trademarks,put the spectrum data of adulterate milk powder samples mixed with the same content of starch of three China trademarks milk powder together to construct a model,the CAR-T of mixed starch content of a sample may reach 0.1%,the CRR-T can reach 1%,if the model construction and predicting were performed with near infrared spectrum data tested in a same day.However,the CAR-T can just stably reach up to 5% and the CRR-T have the same result,if the model construction and predicting were crossly performed with mixed near infrared spectrum data tested in different days.Furthermore,the correct recognizing threshold mixed starch of a sample can stably reach up to 1% and the CAR-T can reach 5%,if the model construction was based on near infrared spectrum data combined the previous four days to predict the output of the another day.On the other hand,we also engaged quantitative analysis to the starch content in milk power with two kinds of arithmetic (PLSR,LS-SVR).In contrast with the testing outputs,the reliability of both the CAR-T and CRR-T in qualitative analysis was further validated.