文献类型：期刊文献

中文题名：基于SWSRA-UVE算法的纸浆材综纤维素近红外预测模型共享研究

英文题名：Research on near infrared prediction model sharing of pulpwood holocellulose based on SWSRA-UVE algorithm

作者：胡云超[1] 王红鸿[1] 熊智新[1] 刘智健[1] 梁龙[2]

第一作者：胡云超

机构：[1]南京林业大学轻工与食品学院,南京210037;[2]中国林业科学研究院林产化学工业研究所,南京210042

年份：2023

卷号：8

期号：2

起止页码：101-108

中文期刊名：林业工程学报

外文期刊名：Journal of Forestry Engineering

收录：CSTPCD;;北大核心:【北大核心2020】；CSCD:【CSCD_E2023_2024】；

基金：中国林科院林业新技术所基本科研业务费专项资金(CAFYBB2019SY039)。

语种：中文

中文关键词：综纤维素;纸浆材;近红外光谱;波长优选;模型传递

外文关键词：holocellulose;pulpwood;near infrared spectroscopy;wavelength optimization;model transfer

分类号：O657.63

摘要：为实现纸浆材综纤维素质量分数的近红外分析模型在3台不同光谱仪器上的共享,提出了光谱信号比值分析筛选波长-无信息变量剔除(SWSRA-UVE)联用算法。即利用无信息变量剔除(UVE)算法,减少光谱信号比值分析筛选波长(SWSRA)方法中无效波长的不利影响,以提高模型传递精度。建立基于SWSRA-UVE算法优选波长的偏最小二乘回归(PLSR)主机模型,并将其模型传递效果与其他常用模型传递方法进行了对比。结果表明,针对2台从机采用SWSRA-UVE方法最终分别优选出的252和105个波长建立模型,分别用于从机1和从机2测量光谱样品的综纤维素质量分数,预测标准偏差(RMSEP)与模型传递前相比分别从2.011 4和9.451 8下降到了1.238 6和1.629 4,且优于SWSRA和其他模型传递算法结果。因此,SWSRA-UVE方法通过UVE算法进一步优选SWSRA一致性波长结果,大大简化了模型传递过程,显著地提高了主机模型的普适性,有利于近红外光谱分析技术的推广应用。
In order to realize the sharing of the near infrared analysis model of pulp holocellulose content among three different spectroscopic instruments, the method of screening wavelengths based on the spectrum ratio analysis(SWSRA) combined with the uninformative variable elimination(UVE) algorithm was proposed. The UVE algorithm was carried out to reduce the adverse effects of invalid wavelengths introduced by the SWSRA method to improve the model transfer accuracy. Using 82 pulp wood flour samples measured on three IAS grating-type near infrared instruments and their holocellulose content data sets, the sample set was divided into a calibration set containing 58 samples and a prediction set containing 24 samples using the Kennard-Stone method. A partial least squares regression(PLSR) master model based on the preferred wavelength of the SWSRA-UVE algorithm was developed, and its model transfer results were compared with those of the SWSRA, segmented direct correction(PDS) and slope intercept(S/B) algorithms alone. The results showed that the models built with 283 wavelengths selected when the UVE method was applied alone were better predicted only for the host samples, but poorer for the slaves, because some of the wavelengths selected by the UVE method from the full spectrum were located in wavelength regions with large differences between different instruments, thus affecting the transfer results of the built models. The model built by the wavelengths selected by the SWSRA method alone can meet the prediction effect of the master and slave samples for practical applications, but the method only considers the selection of wavelengths with small differences between different instruments, which may lead to the inclusion of uninformative and unimportant wavelengths in the wavelengths selected by SWSRA and affect the transfer performance of the SWSRA-PLSR model to some extent. In this study, the models were built for the 252 and 105 wavelengths finally preferred by the SWSRA-UVE method for the two slaves, respectively, and were used to measure the holocellulose content of the spectral samples for slave 1 and slave 2, respectively, and the root mean square error of prediction(RMSEP) decreased from 2.011 4 and 9.451 8 to 1.238 6 and 1.629 4, respectively, and were better than the results of SWSRA alone and other model transfer algorithms. Therefore, the SWSRA-UVE method further optimized the SWSRA consistent wavelength results by the UVE algorithm, which greatly simplified the model transfer process and significantly improved the generalizability of the master model, which was beneficial to the popularization and application of near infrared spectral analysis techniques.