为建立预测能力高、 稳定性强的可见/近红外漫透射光谱无损检测黄花梨可溶性固形物(SSC)数学模型, 对比各种预处理方法、 变量优选方法、 快速独立主成分分析(FICA)以及最小二乘支持向量机(LS-SVM)对黄花梨SSC模型的影响, 得出最佳的组合方法用于建立黄花梨可溶性固形物(SSC)预测模型。 采用Quality Spec型光谱仪采集550~950 nm波段范围内的黄花梨漫透射光谱并采用遗传算法、 连续投影算法和CARS(competitive adaptive reweighted sampling)三种方法筛选黄花梨可溶性固形物的光谱特征变量, 再结合FICA提取光谱主成分, 最后采用LS-SVM建立黄花梨的SSC预测模型。 结果显示, 采用CARS筛选的21个变量, 经FICA挑选出12个主成分数, 联合LS-SVM所建立的CARS-FICA-LS-SVM黄花梨SSC预测模型性能最佳, 建模集和预测集的决定系数及均方根误差分别为0.974, 0.116%和0.918, 0.158%, 同直接采用PLS方法建模相比, 变量数从401个下降到21, 主成分数由14下降到12, 建模集和预测集决定系数分别上升了0.023, 0.019, 而建模和预测均方根误差分别下降了0.042%和0.010%。 CARS-FICA-LS-SVM建立黄花梨SSC预测模型能够有效地简化预测模型并提高预测模型精度。

The purpose of this study was to establish a mathematical model of the visible/near-infrared (Vis/NIR) diffuse transmission spectroscopy with fine stability and precise predictability for the non destructive testing of the soluble solids content of huanghua pear, through comparing the effects of various pretreatment methods, variable optimization method, fast independent principal component analysis (FICA) and least squares support vector machines (LS-SVM) on mathematica model for SSC of huanghua pear, and the best combination of methods to establish model for SSC of huanghua pear was got. Vis/NIR diffuse transmission spectra of huanghua pear were acquired by a Quality Spec spectrometer, three methods including genetic algorithm, successive projections algorithm and competitive adaptive reweighted sampling (CARS) were used firstly to select characteristic variables from spectral data of huanghua pears in the wavelength range of 550~950 nm, and then FICA was used to extract factors from the characteristic variables, finally, validation model for SSC in huanghua pears was built by LS-SVM on the basic of those parameters got above. The results showed that using LS-SVM on the foundation of the 21 variables screened by CARS and the 12 factors selected by FICA, the CARS-FICA-LS-SVM regression model for SSC in huanghua pears was built and performed best, the coefficient of determination and root mean square error of calibration and prediction sets were R2C=0.974, RMSEC=0.116%, R2P=0.918, and RMSEP=0.158% respectively, and compared with the mathematical model which uses PLS as modeling method, the number of variables was down from 401 to 21, the factors were also down from 14 to 12, the coefficient of determination of modeling and prediction sets were up to 0.023 and 0.019 respectively, while the root mean square errors of calibration and prediction sets were reduced by 0.042% and 0.010% respectively. These experimental results showed that using CARS-FICA-LS-SVM to build regression model for the forecast of SSC in huanghua pears can simplify the prediction model and improve the detection precision.