Study On Spectroscopy Method For Identification Of Black Tea Grade And Quality

Chemometrics is an important part of spectral analysis.It includes preprocessing of spectral data,variable selection,and the establishment of qualitative or quantitative models.Each part will affect the final analysis result.This paper takes black tea as the research object,and explores the most suitable variable selection algorithm for the models of black tea grade and its quality with spectrum in range of visible-near infrared spectrum and infrared spectrum as the main research content.Three common variable selection algorithms:successive projections algorithm?SPA?,the competitive adaptive reweighting algorithm?CARS?,the moving window partial least squares method?MWPLS?and moving window smoothing ensemble CARS?MWS-ECARS?algorithm based on the moving window smoothing ensemble strategy are finally chosen in this paper as spectral variable selection algorithms.Based on the previous studies,two improved MWS-ECARS are proposed by changing window smoothing algorithm with Gaussian filter and median filter respectively,and they are named GF-ECARS and MF-ECARS,then we use them to extract characteristic wavelengths of black tea spectrum.In the visible-near infrared spectrum,the partial least square regression?PLSR?models are built with the characteristic wavelengths extracted by the above variable selection algorithms.The regression coefficient?R_P²?of models are compared,the closer the corresponding R_p²is to 1,the better the variable selection algorithm is.In range of the infrared spectrum,the caffeine is used as the representative content of the black tea quality,its theoretical infrared spectrum is used as the evaluation standard.Taking the content of caffeine and the infrared spectrum of black tea as the input variables of four variable selection algorithms,the extracted characteristic wavelengths are compared with the theoretical spectrum respectively,the method whose characteristic wavelengths cover the theoretical spectral information of caffeine most comprehensively is most optimal variable selection algorithm.The results show that:in the visible-near infrared spectral range,the PLSR model built by the characteristic wavelengths extracted by GF-ECARS and black tea grades has good prediction effect,and its R_p² reaches 0.9692.The modeling result of the MF-ECARS algorithm is slightly worse than the original MWS-ECARS,but its R_p² still exceeds 0.96,indicating that the improved algorithms can promote the predictive ability of the original model.The characteristic variables extracted by MWS-ECARS algorithms depend on different window smoothing algorithms.As the width of the smoothing window increases,the continuity of feature variable intervals increases and the number of it decreases.The correlation coefficients of the prediction sets of the three MWS-ECARS algorithms show they are more effective and more stable than three common variable selection algorithms:SPA,CARS and MWPLS.In the infrared spectral range,the GF-ECARS and the original MWS-ECARS algorithm can filter out the wavelengths above 3970cm^-1 that correspond to the O-H bond stretching vibration by setting a frequency threshold.The extracted characteristic wavelengths can more comprehensively cover the spectral information of caffeine than other three variable selections,and they are more convincing from the physical view.In short,the GF-ECARS and MF-ECARS algorithms proposed in this paper have the characteristics of high stability in variable selection and strong continuity of selected characteristic wavelengths.While the stability of the model is improved,the predictive ability is also maintained.It shows that GF-ECARS and MF-ECARS algorithms can provide better spectral data variable selection methods for quantitative or qualitative spectral monitoring models.