连续小波变换结合无参数模型增强框架对时序漂移的近红外光谱建模  

Continuous Wavelet Transform Combined with Parametric-Free Calibration Enhancement Framework for Calibration of Time-shift Near-infrared Spectra

作　　者：张进 叶世著 吴艾璟 李博岩 李杰[3] 詹莜国[3] 彭海根 徐兴阳[3] ZHANG Jin;YE Shi-Zhu;WU Ai-Jing;LI Bo-Yan;LI Jie;ZHAN You-Guo;PENG Hai-Gen;XU Xing-Yang(Key Laboratory of Environmental Pollution Monitoring and Disease Control,Ministry of Education,School of Public Health,Guizhou Medical University,Guiyang 550025,China;Technology Centre,China Tobacco Guizhou Industrial Co.,Ltd,Guiyang 550009,China;Yunnan Tobacco Company Kunming Branch,Kunming 650051,China;Sichuan Vespec Technology Co.Ltd.,Chengdu 610041,China)

机构地区：[1]贵州医科大学公共卫生与健康学院,环境污染与疾病监控教育部重点实验室,贵阳550025 [2]贵州中烟工业有限责任公司技术中心,贵阳550009 [3]云南省烟草公司昆明市公司,昆明650051 [4]四川威斯派克科技有限公司,成都610041

出　　处：《分析化学》2022年第9期1391-1398,共8页Chinese Journal of Analytical Chemistry

基　　金：国家自然科学基金项目(No.22004022);贵州省科技厅科技计划项目(No.黔科合基础-ZK[2021]一般045);贵州省教育厅普通高等学校青年科技人才成长项目(No.黔教合KY字[2021]163)资助。

摘　　要：近红外(Near-infrared,NIR)光谱法具有高效、无损的特点,然而其采集的光谱容易受多种外界因素的影响而发生漂移,导致分析结果出现偏差。时序漂移是一种NIR光谱随检测时间发生持续且有规律漂移的普遍现象。本研究提出了一种时序漂移NIR光谱的建模新方法,将漂移信号分解为背景漂移和样本依赖的时序漂移。分别利用连续小波变换(Continuous wavelet transform,CWT)和半监督-无参数模型增强(Semi-supervised parameter-free calibration enhancement,SS-PFCE)消除NIR光谱中时序背景漂移和样本依赖的时序信号漂移部分,进而实现准确建模。通过对2019年和2020年在云南省境内分别采集的928个和962个土壤样品的时序漂移NIR光谱进行建模,以土壤有机质(Soil organic matter,SOM)含量的预测准确性验证本方法的建模效果。对2019年采集的光谱建模(预测均方根误差(Root mean squared error of prediction,RMSEP)=6.7 g/kg,R 2=0.76),预测2020年采集的漂移光谱时出现了较大的偏差(RMSEP=31.3 g/kg,R 2=0.50)。通过CWT处理后的光谱建模预测,2020年光谱的预测结果明显变好(RMSEP=11.6 g/kg,R 2=0.66);通过SS-PFCE进行模型增强后有了进一步的提升(RMSEP=8.3 g/kg,R 2=0.67)。结果表明,CWT结合SS-PFCE能够最大程度消除NIR光谱中的时序漂移,获得较好的建模结果。Near-infrared(NIR)spectroscopy is a highly efficient and non-destructive technology applied in many areas.The measured spectra,however,are susceptible to shift due to various external interferences,resulting in biased analytical results.Time-shifts in NIR spectra are typical phenomena of continuously shifting spectra with measuring time,which is common in practical applications.In this study,a new method for modeling the time-shift NIR spectra was proposed,in which the shifting signal was divided into shifting background and sample-dependent shifts,and then corrected by continuous wavelet transform(CWT)and semi-supervised parameter-free calibration enhancement(SS-PFCE),respectively.The efficiency of the method was evaluated by modeling 1941 time-shift NIR spectra of soil samples collected in Yunnan Province(China)in 2019 and 2020 to predict the soil organic matter(SOM)content.The spectra taken in 2019 were calibrated(Root mean squared error of prediction(RMSEP)=6.7 g/kg,R 2=0.76)and a large deviation was observed in predicting the spectra taken in 2020(RMSEP=31.3 g/kg,R 2=0.50).After CWT treatment,the prediction of spectra measured in 2020 was significantly promoted by predicting the spectra(RMSEP=11.6 g/kg,R 2=0.66),and the prediction was further improved by SS-PFCE(RMSEP=8.3 g/kg,R 2=0.67).The results indicated that CWT and SS-PFCE were efficient for eliminating the shifting background and sample-related shifts in time-shift NIR spectra,respectively.

关 键 词：近红外光谱 时序漂移 半监督-无参数模型增强 土壤有机质 

分 类 号：O657.33[理学—分析化学]