波长漂移对近红外光谱PLSR分析模型的影响

Effect of Wavelength Drift on PLSR Calibration Model of Near-Infrared Spectroscopy

在近红外光谱分析过程中,单台仪器在不同时间的波长变化及多台仪器间的波长一致与否会对化学计量学定标模型的校正及传递效果产生影响,上述问题可以统一为波长漂移对定标模型的影响。以分析小麦粉中粗蛋白含量为例,首先结合不同谱区光谱数据,利用偏最小二乘回归(PLSR)方法建立了两个定标模型。再由计算机生成不同类型、不同幅度的波长漂移信息,并叠加至验证集样品光谱中,使新光谱相对于定标集光谱产生波长漂移信息。通过考察原定标模型对新光谱的预测与校正情况,研究了波长漂移对PLSR定标模型的影响。结果表明:相对于定标集样品光谱,验证集光谱中无波长漂移信息时,模型的预测标准差(RMSEP)不超过0.3%,预测相关系数不小于0.98;验证集样品光谱在不同波长处的波长漂移信息为一恒定值时,模型的RMSEP会随波长漂移幅度的增大而增大,波长漂移量为-32 cm^(-1)时对应RMSEP为3.69%,预测相关系数变化不大;当验证集样品光谱在不同波长处的波长漂移信息随机变化时,基于长波区光谱所得原始模型的预测结果几乎不受影响;当含有不同波长漂移信息的一系列样品光谱加入到定标集对长波区PLSR分析模型进行校正时,校正后模型的RMSEP为0.3%,几乎不受波长漂移信息的影响,但模型的回归因子数从3显著增大到8,其稳健性变弱;总的来说,当仪器存在波长漂移且幅度不大时,模型预测相关系数几乎不受影响,可通过对预测结果的校正来改善RMSEP,以保证分析结果的准确性。该研究为确定仪器设计参数及分析方法的操作规程,提高近红外光谱分析结果的可靠性提供了实验依据。

Partial least squares regression(PLSR)calibration model will be effect by the wavelength change of a single instrument at a different time and the wavelength consistency of multiple instruments.The process of near-infrared spectroscopy analysis,this problems can be unified as the effect of wavelength drift on chemometric calibration model.In this paper,taking the analysis of crude protein in wheat flour as an example,two calibration models I and II were established by partial least squares regression(PLSR)method within different spectral regions.Different types and amplitudes of wavelength shift information were generated by computer and superimposed into the spectra of the validation set to produce wavelength shift information relative to the spectra of calibration set,the effect of wavelength drift on PLSR calibration model was studied by adding different types and amplitude of wavelength drift information to the spectra of the validation set samples.The results show that the RMSEP of every model is no more than 0.3%and the corresponding Rp is no less than 0.98 when there is no wavelength drift information in the spectra of validation set samples.When the wavelength drift at different wavelengths is constant,the RMSEP increases as the wavelength drift amplitude increases,the RMSEP increases to 3.69%when the wavelength drift is-32 cm^(-1),and the Rp is almost constant;When the wavelength drift varies randomly at different wavelengths,the prediction results of model II based on long wavelength regions are almost not affected.The model II is corrected by a series of spectra added to the calibration set with different wavelength drift information,the RMSEP of the corrected model is 0.3%,The influence of wavelength drift information on RMSEP has been almost eliminated,but the number of regression factors used to establish corredted model increases significantly from 3 to 8,the robustness of the model varies greatly.In general,the RMSEP can be polished by correcting the prediction results to ensure the accuracy of the analysis results if the amplitude of wavelength drift is slight.This study provides an experimental basis for determining the design instrument parameters and operating procedures to improve the reliability of NIR analysis results.