近红外光谱判别分析滤波器的设计与应用  

Design and Application of the Discrimination Filter for Near-Infrared Spectroscopic Analysis

作　　者：孙学辉[1] 赵冰 骆震 孙培健[1] 彭斌[1] 聂聪[1] 邵学广[3] SUN Xue-hui;ZHAO Bing;LUO Zhen;SUN Pei-jian;PENG Bin;NIE Cong;SHAO Xue-guang(Zhengzhou Tobacco Research Institute of China National Tobacco Corporation,Zhengzhou 450001,China;China Tobacco Henan Industry Co.,Ltd.,Zhengzhou 450000,China;Research Center for Analytical Sciences,College of Chemistry,Nankai University,Tianjin 300071,China)

机构地区：[1]中国烟草总公司郑州烟草研究院,河南郑州450001 [2]河南中烟工业有限责任公司,河南郑州450000 [3]南开大学化学学院分析科学研究中心,天津300071

出　　处：《光谱学与光谱分析》2022年第2期399-404,共6页Spectroscopy and Spectral Analysis

基　　金：中国烟草总公司重点实验室项目(110201803001);国家自然科学基金项目(21775076)资助。

摘　　要：近红外光谱(NIRS)在定量和判别分析中已得到广泛应用,化学计量学在其中发挥了重要作用,但仍需要建立基于新原理的方法,简化数据处理和建模过程,使近红外光谱分析更加方便、更加快速。多元光学计算(MOC)技术通过设计合适的光学滤波器可以在光谱测量的同时,根据光谱的整体形状得到定性定量结果。作为一种新的测量和计算方式,近年来在光谱分析领域逐渐得到应用。基于多元光学计算的原理,基于主成分分析和Fisher判别准则设计了近红外光谱的判别滤波器,将近红外光谱投影到二维空间,并在二维空间中计算每一类样品的置信椭圆作为模型进行判别分析。预测样本在二维空间的投影与模型的距离可以作为判别参数,判别值小于等于1时预测样品与模型样品判别为同一类别,否则判别为不同类别,且距离越大,差异性越大。采用460个不同部位的烟叶样品和73个不同生产厂家的药品对所建立的方法进行了测试,表明了方法的准确性。对于三类不同部位烟叶样品和四类不同厂家生产的药品,预测结果的真阳性率可以达到90%以上(除上部烟叶样品外),药品的真阳性率高达95%以上。但烟叶样品的假阳性率仍有些偏高,对于光谱极为相似的实际生产样品结果仍属可接受范围。所建立的方法可推广到其他应用领域,广泛用于基于近红外光谱的质量控制、产品检测、生产一致性监控等。Chemometrics has been widely applied in near-infrared(NIR)spectroscopic analysis for quantitative detection and discrimination.However,new methods are still needed to simplify data processing and modeling to speed up the analysis and improve the convenience in practical uses.As a new type of technique for spectroscopic measurement and computation,the multivariate optical computing(MOC)technique is employed in spectroscopic analysis.The technique uses multivariate information in the spectrum to achieve quantitative computation and discrimination through the designed filters.In this work,the filters for discrimination analysis of near-infrared spectroscopy was designed based on principal component analysis(PCA)and Fisher’s discrimination criterion.The spectra of the calibration samples can be projected into a two-dimensional space by the two filters to achieve an optimized classification,and a confidence ellipse can be obtained for each class of the samples.The ellipse can be used as a model for the discriminating the prediction samples.The distance of a prediction sample to the model is a good measurement of its classification.The samples with a distance less or equal to 1 are classified into the same class of the model,but those with a distance larger than 1 is excluded from the class,and the larger the distance,the bigger the dissimilarity.The proposed method was tested with the NIR spectra of 460 samples of tobacco leaf in three different parts of the plant and 73 samples of the medicinal capsules(amoxicillin granules)produced by four producers.The true positive rate can be higher than 90%,except for the tobacco samples and even higher than 95%for the capsule samples.However,the false-positive rate of the tobacco samples is still not so satisfactory due to the similarity of the NIR spectra.Using near infrared spectroscopy,the proposed method may provide a good technique for quality control,product detection and production monitoring in different fields.

关 键 词：近红外光谱 化学计量学 滤波器设计 置信椭圆 判别分析 

分 类 号：O657.3[理学—分析化学]