Optimized Method of Selecting Samples for Modeling in NIR Spectral Analysis

王丽杰,郭建英,徐可欣

DOI: https://doi.org/10.3969/j.issn.1001-8891.2005.01.018

2005-01-01

Abstract: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.