BiPLS结合GA优选可见/近红外光谱MLR变量

利用反向区间偏最小二乘法(BiPLS)定位光谱糖度若干信息区间,运用遗传算法(GA)从中选择波长点,建立了多元线性回归(MLR)模型.光谱进行卷积平滑和二阶导数处理后,将光谱(225个数据点)分割成25个子区间时,BiPLS优化结果最优.在所定位的信息区间进行GA二次选择特征变量,运行100次依次选择入选频率较高的12个波长点.为简化MLR模型,对于入选的相邻波长选择频率较高者,最后选择638,734,752,868,910,916和938 nm作为同归变量,建立的MLR预测模型相关系数(R2)、校正均方根误差(RMSEC)和预测均方根误差(RMSEP)分别为0.984,0.364和0.471,优于常用的逐步多元线性回归的建模结果.表明BiPLS结合GA可以有效地对李子糖度可见/近红外光谱MLR回归变量进行筛选,提高了模型的精度.

Selection of Variables for MLR in Vis/NIR Spectroscopy Based on BiPLS Combined with GA

The feasibility of using efficient selection of variables in Vis/NIR for a rapid and conclusive determination of fruit inner qualities such as soluble solids content (SSC) of plums was investigated. A new strategy was proposed in the present paper, i. e. two-stage variable selection using the backward interval partial least squares (BiPLS) combined with genetic algorithm (GA). Firstly, it splits the whole spectral region into equidistant sub-regions and then develops all BiPLS regression models, and the informative regions which are used to constructed PLS models with the lowest error can be located. Secondly, GA method is used to select variable in these informative regions, which are used for regression variables of MLR model. The Vis/NIR spectra containing 225 individual data points were processed by Savizky-Golay filter smoothing and second-order derivative, and 9 sub-regions were selected by BiPLS procedure when the spectra were divided into 25 sub-regions. The optimal 12 variables, which were the output of the GA procedure, were selected by the higher occurrence frequency while the GA procedure ran 100 times. In order to simplify the multiple linear regression (MLR) modeling, the wavelength variables with the maximum occurrence frequency were chosen when the adjacent wavelengths were selected by GA. Finally, 638, 734, 752, 868, 910, 916 and 938 nm were used to build a MLR model. The results show that MLR model produced by BiPLS-GA performs well with correlation coefficients (R) of 0. 984, root mean standard error of calibration (RMSEC) of 0. 364 and root mean standard error of prediction (RMSEP) of 0. 471 for SSC, which outperforms models using stepwise regression analysis (SRA). This work proved that the BiPLS-GA could determine optimal variables in Vis/NIR spectra and improve the accuracy of model.