液态样本近红外光谱测量中的光程变化误差消减方法研究

以蔗糖溶液为研究对象，利用近红外光谱分别测量4，5和6 mm光程下不同浓度蔗糖溶液的透反射光谱，研究采用矢量归一化、基线偏移校正、多元散射校正、标准正态变量变换、一阶导数5种预处理方法消除光程差异的影响，并结合PLS方法建立校正集模型。与原始光谱的PLS模型相比，五种预处理方法均对模型的预测精度有不同程度的提高，其中，多元散射校正结合PLS方法建立的模型最优，使原始光谱的主成分数PC由6下降为3，决定系数R2由0.891 278提高到0.987 535，交互验证决定系数R2_CV由0.888 374提高到0.983 343，校正标准偏差RMSEC由1.704%下降到0.89%，交互验证的校正标准偏差RMSECV由1.827%下降到1.05%，预测集样本的相关系数由0.950 89上升到0.976 22，预测标准偏差由0.014 36下降为0.01。结果表明，五种预处理方法中，多元散射校正法能够消除光程差异的干扰，提高模型的预测精度，改善稳定性。

Research on Error Reduction of Path Change of Liquid Samples Based on Near Infrared Trans-Reflective Spectra Measurement

Based on sucrose solution as the research object, this paper measured the trans-reflective spectrum of sucrose solution of different concentration by the technique of near infrared spectrum in three optical path (4, 5, 6 mm). Five kinds of pretreatment method (vector normalization, baseline offset correction, multiplicative scatter correction, standard normal variate transformation, a derivative) were used to eliminate the influence of the optical path difference, and to establish model of the calibration set in combination with the PLS(Partial Least Squares)method. Five kinds of pretreatment method could restrain the interference of light path in varying degrees. Compared with the PLS model of original spectra, the model of multiple scattering correction combined with PLS method is the optimal model. The results of quantitative analysis of original spectra: the number of principal component PC=6, the determination coefficient R2=0.891 278, the determination coefficient of cross validation R2_CV=0.888 374, root mean square error of calibration RMSEC=1.704%, root mean square error of cross validation RMSECV=1.827%; The results of quantitative analysis of spectra after MSC pretreatment: the number of principal component PC=3, the determination coefficient R2=0.987 535, the determination coefficient of cross validation R2_CV=0.983 343, root mean square error of calibration RMSEC=0.89%, root mean square error of cross validation RMSECV=1.05%. The correlation coefficient of the prediction set is as much as 0.976 22. root mean square error of prediction is 0.01, lesser than 0.014 36. The results show that the MSC can eliminate the influence of optical path difference, improve the prediction precision and improve the stability.