Study Of Terahertz Spectra With Factor And Independent Component Analysis

Terahertz spectroscopy has a wide range of applications in the analysis and detection of substances because terahertz waves possess many advantages such as low energy,broad spectrum,high penetration,and specific absorption.On the other hand,machine learning methods have the advantages of no human intervention,automation,and scalability.With the impetus of machine learning,the data analysis capability and application scope in terahertz spectrum technology have been significantly improved.However,the application of machine learning is limited when dealing with unlabeled terahertz spectra data.In this thesis,the factor analysis and independent component analysis methods in unsupervised machine learning are extended and applied to the analysis of unlabeled terahertz spectra data,and three terahertz spectrum analysis methods are constructed and applied to compound classification and identification,mixture quantification,and contaminant detection with excellent results.First,the factor analysis method is extended and applied to the terahertz spectrum analysis,and the terahertz factor analysis method is constructed.For the terahertz spectra of different molecules,the determination of molecular species is accomplished using the factor analysis method when the information of molecular species is completely unknown in the modeling process,and the classification of some species samples could reach 90%accuracy.The common factors of these molecules are also obtained,providing information for the determination of molecular structures,and can be used for the classification and identification of new samples.This provides a new method for the classification and identification of unknown samples,especially when unlabeled data and small sample size limits the use of commonly used supervised learning methods.Secondly,the independent component analysis method is extended and applied to the analysis of terahertz spectra to construct the terahertz independent component analysis method.In experiment,mixtures of milk powder and two different additives(glucose and melamine)are used as the samples.While the composition of the mixtures is unknown during the modeling process,the use of independent component analysis is effective in obtaining information of the composition and content of the mixtures with relative errors of less than 18% for the component content.Terahertz spectra of each component are also obtained,which are in good agreement with the spectra directly measured and could further provide a basis for the identification of the species of the mixtures' components.This provides a new solution for the analysis of unknown mixtures,where the spectral data can be used to separate and quantify components directly,instead of separating the sample components and then testing them in sequence.Finally,the factor analysis method and independent component analysis method are applied to the terahertz spectrum analysis,and the terahertz factor and independent component integrated analysis method is constructed.In experiment,the total suspended particulate matter(TSP)in the air pollutants in Wuhan area is collected and measured.The pollutant particle samples and their terahertz spectra are obtained after 128 days and 4months of observation.Without physical separation of pollutants,the absorption spectra of the four main components are obtained by independent component analysis using only the directly measured data of the samples,i.e.,the terahertz absorption matrix,and then combined with factor analysis to determine the species of the four components.After comparison,it is found that the types and contents of the components obtained experimentally are consistent with official data and data reported in the literature.This provides a more direct,rapid,and simple analytical method for the detection of pollutants.