基于机器学习和中红外光谱的牦牛奶掺假预测模型研究  

Prediction adulteration of yak milk based on machine learning and mid-infrared spectroscopy

作　　者：褚楚 罗雪路 王海童 温佩佩 杜超 丁考仁青 拉毛草 张淑君 CHU Chu;LUO Xuelu;WANG Haitong;WEN Peipei;DU Chao;Dingkaorenqing;Lamaocao;ZHANG Shujun(College of Animal Science and Technology,College of Veterinary Medicine,Huazhong Agricultural University,Wuhan 430070,China;Animal Husbandry Station of Gannan Tibetan Autonomous Prefecture,Hezuo 747000,China;Animal Disease Prevention and Control Center of Luqu County,Gannan Prefecture,Luqu 747200,China)

机构地区：[1]华中农业大学动物科学技术学院、动物医学院,武汉430070 [2]甘南藏族自治州畜牧工作站,合作747000 [3]甘肃省碌曲县动物疫病预防控制中心,碌曲747200

出　　处：《华中农业大学学报》2025年第2期116-124,共9页Journal of Huazhong Agricultural University

基　　金：国家重点研发计划-政府间国际科技创新合作(2021YFE0115500);湖北省国际合作项目(2022EHB043)。

摘　　要：为监督和规范牦牛奶生产和销售,对牦牛奶中奶牛奶的掺假比例进行定量预测,研发新的快速检测技术,通过中红外光谱技术结合机器学习算法建立检测牦牛奶中掺加奶牛奶的预测模型,以76份纯牦牛奶、76份掺加10%奶牛奶的牦牛奶、76份掺加25%奶牛奶的牦牛奶、76份掺加50%奶牛奶的牦牛奶为研究对象,利用5种光谱预处理算法、6种定性和12种定量机器学习算法,分别建立鉴别纯牦牛奶和掺加奶牛奶的牦牛奶的二分类定性模型和预测掺加奶牛奶比例的定量回归模型。结果显示,基于支持向量机建模算法、无预处理光谱建立的鉴定纯牦牛奶和掺加奶牛奶的牦牛奶的预测模型效果最好,该模型验证集AUC为0.95,准确性0.84,灵敏度0.93,特异性0.87,可用于纯奶和掺假奶的鉴定。利用贝叶斯正则化神经网络建模算法和一阶导数光谱预处理算法建立了预测牦牛奶中奶牛奶掺加比例的最佳定量模型,该模型R_(P)^(2)=0.88,RMSEV=6.57%,RPD=2.89%。结果表明,中红外光谱技术结合机器学习算法可有效地鉴定出掺加奶牛奶的牦牛奶,并可检测出掺假的比例。A predictive model for detecting the addition of milk to yak milk was established by combining mid-infrared spectroscopy(MIRS) with machine learning algorithms to supervise and regulate the production and sale of yak milk,further quantitatively predict the proportion of adulteration in yak milk and provide new technology of rapid detection.76 samples of pure yak milk,76 samples of yak milk adulterated with 10% milk,76 samples of yak milk adulterated with 25% milk,and 76 samples of yak milk adulterated with 50% milk were used to establish binary qualitative models for distinguishing pure yak milk from yak milk adulterated with milk,and quantitative regression models for predicting the proportion of yak milk adulterated with milk with five spectral preprocessing algorithms,six qualitative and twelve quantitative machine learning algorithms.The results showed that the predictive model for identifying pure yak milk and yak milk adulterated with milk based on support vector machine modeling algorithm and the spectrum without preprocessing had the best performance.The validation set AUC,the accuracy,the sensitivity,and the specificity of the model was 0.95,0.84,0.93,and 0.87,which can be used for the identification of pure milk and adulterated milk.The optimal quantitative model for predicting the proportion of milk in yak milk was established using Bayesian regularized neural network modeling algorithm and first-order derivative spectral preprocessing algorithm.The model had R_(P)^(2)=0.88,RMSEV=6.57%,and RPD=2.89%.It is indicated that the combination of mid-infrared spectroscopy and machine learning algorithms can effectively identify yak milk adulterated with milk and detect the proportion of adulteration.

关 键 词：中红外光谱(MIRS) 机器学习 牦牛奶 牛奶掺假 预测模型 

分 类 号：O657.33[理学—分析化学] TS252.7[理学—化学]