论文部分内容阅读
为提高校正模型的预测精度,以烟草中淀粉近红外光谱(NIR)校正模型为研究对象,分别利用全光谱波段(FS)、方差光谱(VS)筛选光谱变量和遗传算法(GA)筛选光谱波长,结合偏最小二乘法建立校正模型(FS+PLS、VS+PLS和GA+PLS),并对100个初烤烟叶样品进行了预测。结果显示:1FS+PLS(变量数1 557个)、VS+PLS(变量数781个)和GA+PLS(变量数72个)3种校正模型的决定系数Rc2、交互验证均方根误差(RMSECV)分别为0.976 4、0.433,0.987 1、0.332和0.988 5、0.314。2与FS+PLS和VS+PLS模型相比,GA+PLS模型的光谱变量数分别减少为FS和VS变量数的4.62%和9.22%,主因子数由15降至12,Rc2明显提高,RMSECV显著降低。3FS+PLS、VS+PLS和GA+PLS模型对100个初烤烟叶样品的预测结果显示,Rp2、预测均方根误差(RMSEP)分别为0.965 2、0.780,0.984 3、0.501和0.985 3、0.496,预测值与其对应的化学检测值之间通过配对T检验,显著性Sig.值、T值和平均相对误差(%)分别为0.271、1.107、17.48%,0.973、0.034、13.13%和0.722、0.357、13.12%,3种方法所建立校正模型的预测值与检测值之间均无显著性差异,模型预测精度(RSD)分别为10.34%、6.98%和4.76%。基于逐步优化光谱信息法建立的GA+PLS校正模型的预测精度优于FS+PLS和VS+PLS模型,该方法对于提高复杂化学体系模型的精度有参考意义。
In order to improve the prediction accuracy of the calibration model, the model of starch NIR correction in tobacco was selected as the research object. Spectral variables (GA) and genetic algorithm (GA) were used to screen spectral wavelengths (FS + PLS, VS + PLS and GA + PLS) were combined with partial least squares to predict 100 samples of early-cured tobacco leaves. The results showed that the coefficient of determination (Rc2) of three calibration models: 1FS + PLS (variable number 1 557), VS + PLS (variable number 781) and GA + PLS (variable number 72) ) Were 0.976 4,0.433,0.987 1,0.332 and 0.988 5,0.314.2 Compared with the FS + PLS and VS + PLS models, the number of spectral variables in the GA + PLS model was reduced to 4.62% of the number of FS and VS variables, respectively And 9.22% respectively. The number of major factors decreased from 15 to 12, Rc2 significantly increased and RMSECV significantly decreased. The predicted results of 100 samples of pre-cured tobacco leaves from 3FS + PLS, VS + PLS and GA + PLS models showed that the root mean square error of prediction (RMSEP) of Rp2 was 0.965, 2.0.780, 0.984 3, 0.501 and 0.985 3, , The predicted value and the corresponding chemical test value were tested by paired T test. The significant Sig. Value, T value and average relative error (%) were 0.271, 1.107, 17.48%, 0.973, 0.034, 13.13% and 0.722, 0.357 , 13.12% respectively. There was no significant difference between the predicted value and the measured value of the calibration model established by the three methods. The model predictive accuracy (RSD) was 10.34%, 6.98% and 4.76% respectively. The prediction accuracy of GA + PLS calibration model based on step by step optimization of spectral information method is superior to that of FS + PLS and VS + PLS models. This method has reference value for improving the accuracy of complex chemical system models.