Study On Fuzzy Discriminant Information Extraction Of Near Infrared Spectra Of Tea

With the pursuit of high-quality life and the fast-developing of modern food industry,the economic benefits of tea industry have been gradually improved,but the problem of fake and shoddy products has become increasingly prominent.Therefore,the realization of rapid and accurate identification of tea quality is particularly important.At present,near infrared(NIR)spectroscopy,as an emerging detection technique,is rapidly popularized in the field of food detection with its unique advantages.However,NIR spectral data have some characteristics: highdimensional,overlapping,redundant and nonlinear,which cause a series of technical difficulties,such as dimension disaster and information extraction difficulties,when features from them are extracted for better classification.Hence,this paper focuses on the feature extraction of NIR spectral data in the qualitative modeling of tea.Aiming at the conundrum that the “hard” feature extraction algorithm is difficult to describe the information diversity of sample class,and the limitation that the linear feature extraction algorithms cannot to deal with the problem of linear inseparability,the fuzzy set theory and kernel function theory are used to establish the feature extraction model of the NIR spectra of tea.The main contents and conclusions are as follows:(1)Qualitative modeling of tea NIR spectra based on the linear feature extraction algorithms.Firstly,four kinds of “hard” linear extraction algorithms,namely principal component analysis(PCA),linear discriminant analysis(LDA),uncorrelated linear discriminant analysis(ULDA)and orthogonal linear discriminant analysis(OLDA),were studied to obtain effective identification information of NIR spectral data of tea.Secondly,considering the shortcoming of “hard” feature extraction methods in describing the diversity of sample information,two “soft” identification information extraction algorithms were proposed by introducing the fuzzy set theory to the existing “hard” feature extraction algorithms,including fuzzy uncorrelated linear discriminant analysis(FULDA)and fuzzy orthogonal linear discriminant analysis(FOLDA).Finally,the classification results of tea NIR spectra qualitative analysis models based on the above linear feature extraction algorithms were discussed.The results showed that the recognition rates of classification models based on supervised feature extraction algorithms were better than those based on unsupervised feature extraction algorithms.At the same time,the classification accuracies of the models based on the two proposed fuzzy feature extraction algorithms were higher than 95.93%,which verified the feasibility of the “soft” identification information extraction algorithms for extracting the effective classification features from the NIR spectral data of tea.(2)Feature extraction model of NIR spectra based on the kernel trick.The NIR spectral data of samples usually contain some nonlinear features,and the linear feature extraction algorithms lacks the ability to capture the nonlinear discriminant information of data.In this paper,by introducing the kernel function theory into the existing linear feature extraction algorithms,a kernel-based fuzzy orthogonal linear discriminant information(KFOLDA)feature extraction algorithm was proposed.In addition,the proposed KFOLDA algorithm was used to establish the classification model of four brands of tea in Anhui,and the model was compared with that of kernel-based orthogonal linear discriminant analysis(KOLDA)algorithm.The results showed that the recognition rates of the tea classification models based on KOLDA or KFOLDA can reach 100%,and the classification effect of the model based on linear kernel function was the best,which indicated the effectiveness of the proposed KFOLDA feature extraction algorithm in obtaining the nonlinear discriminant information of the data.