摘要
结合牛奶成分近红外光谱测量系统的实例,在已定的浓度范围内针对牛奶中脂肪、蛋白质、乳 糖三成分采用正交设计法优选参与建模的样品。研究中首次利用正交表的"正交性"原理优选建模样 品,并针对牛奶中脂肪浓度的测量采用偏最小二乘(PLS)回归方法交互验证方式建立模型。在此基 础上,将正交设计样品集与常规方法选择的样品集的脂肪PLS模型的预测结果进行了对比。实验结果 表明:采用正交设计样品集与常规样品集分别建立的PLS模型的预测偏差之差低于0.02g/100g,上述 两种方法PLS模型的实际预测浓度与参考浓度之差均集中在0.1g/100g,而后者样品数量约为前者的七 倍。进一步的实验结果表明:从常规样品集的样品中随机抽取与正交设计样品集的样品数量相同的样 品作为随机样品集并建模,其PLS模型的预测偏差高于常规方法的两倍、相关系数相对较低,并且其 实际预测浓度与参考浓度之差集中在0.4g/100g。
In order to reduce excessive experiments and to improve model applicability, for the first time, the samples for modeling are selected by the Orthogonal Design Method and applied to NIR Spectral System for measuring milk constituents. By selecting the samples for modeling using the principle of orthogonality in the orthogonal table to fat, protein and lactose of milk, Partial Least Square (PLS) Regression model was built by means of cross validation for measuring the fat concentration of these samples, and the predictions of this model and other two models are compared. To the latter two models, a model built with samples selected by the conventional method, the other model whose sample set size is the same as the orthogonal sample set built with samples extracted randomly from the samples already selected by the conventional method. The results indicate that the difference between the prediction error of the orthogonal model and that of the conventional model is less than 0.02g/100g. For these two models, the discrepancy between the predicted concentration and the reference concentration is about 0.1g/100g. However, the sample size of the orthogonal model is about seven times that of the conventional model. Furthermore, the prediction error of the third model is larger than that of the former two models, and its correlation coefficient is smaller than the correlation coefficients of the former two models. In addition, the difference between the predicted concentration from the third model and reference concentration is about 0.4g/100g.
出处
《红外技术》
CSCD
北大核心
2005年第1期75-78,共4页
Infrared Technology
关键词
PLS
脂肪
正交设计
首次
样品
预测
牛奶成分
浓度
数量
近红外光谱分析
Near-Infrared (NIR) Spectral Analysis
Orthogonal Design Method
Orthogonality
Milk
Partial Least Square (PLS) Regression
