基于尺度不变特征变换筛选稳定特征波长的近红外光谱模型传递方法

A Near Infrared Spectroscopy Calibration Model Transfer Method Based on Scale Invariant Feature Transform to Select Stable Characteristic Wavelengths

该研究利用一维尺度不变特征变换(SIFT)算法寻找烟叶近红外光谱(Near infrared spectroscopy,NIRS)的稳定特征波长,根据样品精密度测试光谱筛选的波长计算重现率和重现度,采用L9(33)正交表优化SIFT算法中的相关参数,使重现率和重现度尽可能高。基于优化的参数和主机上10个代表性样品的光谱,筛选出10个稳定特征波长集合,以这些波长集合并集的光谱响应为自变量,采用偏最小二乘(PLS)方法构建烟叶总植物碱NIRS模型(简称SIFT-PLS)。该模型直接传递到3台从机后,对3台从机样品总植物碱的平均相对预测误差(MRE)均满足小于6%的企业内控要求,而全光谱模型(WW-PLS)直接转移后仅1台从机的MRE满足要求,经分段直接校正(PDS)方法校正从机光谱后,WW-PLS模型也仅对1台从机的MRE小于6%。采用SIFT算法筛选稳定特征波长建立的NIRS模型可在3台从机直接共享,无需转移集,不需对从机光谱或光谱模型进行校正,实现了真正意义的无标样NIRS模型的直接转移。

In this study,one-dimensional scale-invariant feature transform(SIFT)algorithm was used to find stable characteristic wavelengths of the near infrared spectra(NIRS)of tobacco leaves.The reproducibility rate and reproducibility were calculated according to the wavelengths selected by SIFT based on the precision tests of spectra.The orthogonal table of L9(33)was adopted to optimize the parameters in the SIFT algorithm to make the reproducibility rate and reproducibility as high as possible.Based on the optimized parameters and the spectra of 10 representative samples tested on the master instrument,10 stable characteristic wavelength sets were selected,and the spectral responses of the union of these wavelength sets were used as the independent variables to build a NIRS model for predicting the total alkaloids in tobacco leaves by partial least squares(PLS)method,which was shortened as SIFT-PLS model.After the model was directly transferred to three slave instruments,the average relative prediction errors(MREs)of the total alkaloids in tobacco leave samples tested on the three slaves were all smaller than 6%,meeting the internal control requirements of tobacco enterprises.Nevertheless,when the whole wavelength model(WW-PLS)was directly transferred to the three slaves,MRE of only one slave instrument satisfied the requirements.Even the spectra of the samples tested on the three salves were corrected by piecewise direct standardization(PDS)method,MRE of samples tested on only one slave was lower than 6%.The robust SIFT-PLS near infrared spectroscopy model established by selecting stable characteristic wavelengths with SIFT method could be directly shared by the three slave instruments,without needing transfer sets,and correcting spectra of slave instruments or spectral calibration models.The method proposed in this paper could be used to realize the NIRS calibration model transfer without standards in the true sense.