基于遗传算法的苹果糖度近红外光谱分析

Analysis of Near Infrared Spectra of Apple SSC by Genetic Algorithm Optimization

遗传算法(GA)应用在偏最小二乘法(PLS)校正模型的波段优化选择中具有显著的效果。将遗传算法用于波段选择,能更快达到最优解,有效提高测量精度,减少建模所用变量。文章研究了在近红外苹果糖度无损检测中,遗传算法作为模块进行波段选择,建立了GA-PLS模型;为了说明遗传算法优选波段可行性,另外建立了全谱和经验谱区的PLS定量模型,并评价了模型的稳健性。首先对傅里叶变换近红外光谱进行多元散射校正、Savitky-Golay卷积平滑后,用遗传算法优选波段(R-SGA),参与建模数据点从原始1550减少到434个。然后采用一阶导数光谱建立GA-PLS模型,相比全谱PLS(1550个数据点)和经验谱区PLS(717个数据点)模型具有更高的预测精度,其建模结果为RC=0.966,RMSEC=0.469,RP=0.954,RMSEP=0.797。结果表明,遗传算法可用于PLS法建立苹果糖度校正模型前的数据优化筛选,有效提高测量精度,并减少建模变量。

Genetic algorithm （GA） is an effective method in regions selection applied in building multivariate calibration model based on partial least squares regression. If genetic algorithm is run repeatedly as a block, the optimal solution is obtained faster, the numbers of data used to build calibration model are further reduced, and the prediction precision is further improved. An efficient method named region selecting by genetic algorithms （R-SGA） for building a PLS calibration model of NIR is presented in the present paper, in which each gene of chromosome represents a sub-region. In the R-SC-A method, one needs to divide aver- agely the full spectral band into many sub-regions, and to build a research space with all the combinations of these sub-regions. The FT-NIR spectra were processed by GA after MSC and Savitky-Golay smoothing, a PLS calibration model of NIR was built by using the optimal combinations of these sub-regions. Meanwhile, the full region selecting PLS （FS-PLS） and experiential region selecting PLS （ES-PLS） models were developed using spectra after first-order derivative pretreatment. The seven intervals selected by region selecting by R-SGA which contained 434 variables were used as calibration set in GA-PLS. The prediction precision of GA-PLS model was better than FS-PLS and ES-PLS models, with Rc =0. 966, RMSEC=0. 469, Rv=0. 954 and RM- SEP=0. 797. It was concluded that by using GA technique, in the pretreatment of apple SSC model by PLS, it is possible to optimize data selecting, enhance the precision of prediction and reduce the number of variables of calibration.