Research On Tea Quality Analysis Based On Infrared Spectroscopy And Raman Spectroscopy

Tea is one of the most popular drinks in the world.It is not only beneficial to human health,but also has high economic value.The market price of different varieties and grades of tea varies greatly.The rapid and accurate identification of tea varieties and grades is of great significance to promote the development of China’s tea industry.The traditional identification methods mainly include sensory evaluation method and physical and chemical analysis method.Sensory evaluation method relies on manual evaluation experience,with strong subjectivity in the evaluation process and lack of accuracy in the results.Although the physical and chemical analysis method has high precision and accurate results,it has long test cycle,complex instrument operation and high test cost.Infrared spectroscopy has the advantages of fast,nondestructive and pollution-free.It has a good application prospect in the identification of tea varieties and grades.In this paper,the identification model of tea varieties and grades is established based on near-infrared spectroscopy technology.Aiming at the problems existing in infrared spectroscopy,such as low accuracy of model recognition,poor model mobility and generalization,high cost of collecting a large number of spectral data,large dimension of spectral data and so on.In addition,pesticide residues are one of the important factors affecting the quality of tea.Traditional mass spectrometry,chromatography and other detection methods have some problems,such as complex sample processing and instrument operation,and the detection cost is high and the time is long.Generally,it can only be completed in the laboratory,so it is difficult to detect pesticide residues in real time.Due to the limitation of instrument sensitivity,infrared spectroscopy is difficult to realize the trace detection of pesticide residues in tea.Surface enhanced Raman spectroscopy(SERS)has the advantages of high sensitivity and fast detection speed,which can make up for the deficiency of infrared spectroscopy in the detection of pesticide residues in tea.Taking tea samples containing carbendazim pesticide residues as the research object,the trace detection of carbendazim residues in tea was realized by surface enhanced Raman spectroscopy.The main research contents and results of this paper are as follows:(1)The identification of tea varieties and grades is realized based on feature selection and convolution neural network.Firstly,the wavelet transform algorithm(WT)is used to denoise the near-infrared spectral data of five varieties and four grades of tea,and then the features of the denoised spectral data are selected.The dimension reduction of spectral data is realized by using joint interval partial least squares(sipls),continuous projection algorithm(SPA)and competitive adaptive reweighting algorithm(cars),which solves the problem of large dimension of spectral data.Finally,a convolutional neural network(CNN)classification model is established for the spectral data after feature selection to realize the identification of tea varieties and grades.The results showed that in the identification of tea varieties and grades,the accuracy of CNN model was only 66.7%and 75%,spa-cnn model reached 95.83%and 96.67%,and cars-cnn further improved the accuracy to 97.72%and 98.67%.(2)Identification of tea varieties based on dcgan extended near infrared spectroscopy data set.In the identification of tea varieties,the collection of near-infrared spectrum data is restricted by factors such as manpower,instrument and cost.It is difficult to collect enough data,which affects the detection accuracy.Taking eight different varieties of tea as the research object,the nearinfrared spectrum data set is expanded by translation method,linear superposition method,adding noise method and deep convolution generation countermeasure network(dcgan).The results show that the data set expanded by dcgan can significantly improve the accuracy of tea variety identification.On this basis,the number of iterations of the samples generated by dcgan is studied.When the number of iterations is 6000,the quality of the generated spectrum is better.Different classification models are used to model and analyze the extended spectral data.The experimental results show that the accuracy of SVM model is improved from 67.45%to 90.46%,gbdt model is improved from 68.75%to 90.42%,xgboost model is improved from 65.6%to 88.83%,CNN model has the highest accuracy,and the model accuracy is improved from 73.36%to 98.68%.Dcgan can improve the accuracy of tea variety identification,solve the problem of high cost of spectral data acquisition,and reduce the complexity of modeling.(3)Research on the migration ability and generalization ability of pca-cnn model.In order to solve the problem of poor transfer of traditional models between different spectral data sets and different instruments,pca-cnn model was established for the spectral data of 8 kinds of tea expanded by dcgan in near infrared band.Fix the structural parameters of the first three layers of convolution layer,fine tune the parameters of the full connection layer,and migrate the pca-cnn model to the spectral data of tea varieties and grades in the mid infrared and terahertz bands.The results show that in the spectral data of mid infrared tea varieties and grades,through re modeling,pca-cnn model can achieve 99.43%and 100%accuracy,and the accuracy of model transfer learning is 98.77%and 99.43%.In the data of tea varieties and grades in terahertz band,the accuracy of re modeling is 95.67%and 96.45%,and the accuracy of transfer learning is 95.32%and 95.76%respectively.Based on pca-cnn transfer learning model,the accuracy similar to that of re modeling can be achieved.Secondly,several models such as pca-rf,PCA-SVM,pca-knn and PCA xgboost are compared.Pca-cnn model has good migration ability.In the open source olive oil origin identification data set and drug data set tested by different instruments(idrc2002),pca-cnn model achieves high accuracy.In addition,different intensities of random noise are added to the near-infrared spectrum data set based on dcgan.Compared with other models,CNN has better anti noise ability.When 60dB random noise is added,the accuracy of CNN model can reach 92%.Therefore,convolutional neural network has stronger generalization ability in realizing tea variety and grade identification.(4)Trace detection of carbendazim residues in tea based on SERS.It is a kind of fungicide that can easily penetrate the leaves of tea.The detection of SERS is disturbed by pigments,tea polyphenols and other organic substances in tea.Tea samples must be extracted and purified before treatment.In this paper,the tea samples containing carbendazim residues were studied.Firstly,the tea samples containing pesticide residues were pretreated to prepare 60NM silver nanoparticles as surface reinforcing substrate.The experimental results showed that the quantitative detection model of carbendazim residue in tea was established by least square regression method,y=1199x+1021,and the correlation coefficient was 0.9936.The minimum detection concentration of this method is lmg/L,which meets the requirements of 5mg/L specified in the national standard.In addition,comparative experiments based on mid infrared spectroscopy and liquid chromatography tandem mass spectrometry(LC-MS)were designed to detect carbendazim residues in tea.Through the comparative analysis of the experimental results,the minimum detection concentration and detection accuracy of carbendazim residue based on SERS are better than that of mid infrared spectroscopy.The analysis results of SERS technology are basically consistent with those of liquid chromatography tandem mass spectrometry,but the experimental process is simpler.Therefore,SERS technology can quickly and accurately detect carbendazim in tea.To sum up,the method proposed in this paper can well solve the common problems of spectral technology in the study of tea quality,such as low accuracy of multi classification model,high cost of spectral data collection of a large number of samples,weak model migration ability and generalization ability.Secondly,it is verified that the method proposed in this paper can be applied to spectral detection in other fields and has strong universality.In addition,a method for the determination of carbendazim in tea was proposed,which provides a reliable method for the determination of other pesticide residues in tea.