PLS-GRNN法近红外光谱多组分定量分析研究

研究了偏最小二乘(partial least squares,PLS)与广义回归神经网络(generalized regression neural networks,GRNN)联用在近红外光谱多组分定量分析中的应用.以饲料样品为实验材料,采用PLS-GRNN法建立了饲料中水溶性氯化物、粗纤维、脂肪三项组分含量近红外光谱定量分析模型.马氏距离法剔除强影响点和奇异点,用PLS法将原始数据压缩为主成分,取8个主成分吸收峰与4个原始图谱特征峰值输入GRNN网络,网络光滑因子σi为0.1.PLS-GRNN模型对样品3个组分含量的预测决定系数(r2)分别为:0.984 0,0.987 0,0.983 0;样品平行扫描光谱预测值的标准偏差分别为:0.003 26,0.065 5,0.031 4.结果表明所建PLS-GRNN模型通过近红外光谱能够准确预测饲料中水溶性氯化物、粗纤维、脂肪三项组分含量,为近红外光谱进行多组分定量分析提供了新思路,同时为解决近红外快速检测技术在预测组分含量较低的样品时误差相对较大的问题提供了可靠的方法.

Multicomponent Quantitative Analysis Using Near Infrared Spectroscopy by Building PLS-GRNN Model

The present paper introduces an application of near infrared spectroscopy(NIRS) multi-component quantitative analysis by building partial least squares (PLS)-generalized regression neural networks (GRNN) model. The PLS-GRNN prediction model for chlorine, fibre and fat in 45 feedstuff samples was established with good veracity and recurrence. Eight peak values in principal components compressed from original data by PLS and four in original spectra were taken as inputs of GRNN while 4 predictive targets as outputs. 0.1 was chosen as smoothing factor for its good approximation and prediction with the lowest error compared with 0.2, 0.3, 0.4 and 0.5. Predictive correlation coefficient and Standard error of the estimate of three components by the model are 0.984 0, 0.987 0 and 0.983 0, and 0.015 89, 0.154 1 and 0.115 1, while the Standard deviations of an unknown sample scanned 8 times are 0.003 26, 0.065 5 and 0.031 4. The results show that PLS-GRNN used in NIRS is a rapid, effective means for measuring chlorine, fibre in the fat in feedstuff powder, and can also be used in quantitative analysis of other samples. A settlement in the high error of prediction of other samples with lower contents was also shown.