正交信号校正用于偏最小二乘建模过拟合现象的研究

摘要

通过用近红外光谱和PLS建立测定二元、三元调合食用油中花生油含最模型以及二甲亚砜水溶液浓度模型,使用交互验证统计参数,包括决定系数RC,标准偏差SEC,预测值和实际值线性拟合方程的斜率a和截距b,和外部验证统计参数,包括决定系数RV,标准偏差SEP,预测值和实际值线性拟合方程的斜率a和截距b,评价建模效果,比较了分别使用原始光谱和正交信号校正(OSC)处理光谱的PLS建模结果.研究结果表明:OSC校正可明显改善近红外光谱与被测物质浓度的线性相关性.虽然OSC预处理可以改善PLS模型的交互验证结果,似是外部验证结果变差,即正交信号校正用于PLS建模将导致过拟合现象.通过分析算法原理,认为OSC与PLS1对光谱中与性质无关的信息作了重复剔除,误删除了部分有效信息,导致过拟合效应.%In the present paper, the over-fitting phenomenon in building PLS model using orthogonal signal correction (OSC) was studied through establishment of quantitative calibration models for the peanut oil content in blending edible oils, and for the dimethylsulfoxide concentration in water solution. The cross validation results and the predication results of PLS models using OSC and without using OSC were compared to evaluate the effectiveness of OSC for improving the performance of PLSl model.The results show that the application of OSC to PLS modeling will lead to an over-fitting phenomenon. According to the principles of their algorithms, when OSC and PLS are used together, the signals which are not correlated to the interested property are removed twice from the raw spectra. This leads to deleting the parts of useful information in spectra, and to spoiling the predictive ability of PLS models to some extent.