The Quality Evaluation System For Tieguanyin Tea Based On Chemical Analysis And Machine Learning

Tieguanyin is one of the top ten famous teas in China for its unique flavor and healthy functions.The existing quality evaluation methods of Tieguanyin have severely restricted the benign development of its industry.The current quality evaluation methods of Tieguanyin tea mainly depend on the manual evaluation,which is susceptible to subjective factors.So,the presented quality evaluation methods restrict the development of the production process and establishment of Tieguanyin tea brand.Machine learning(ML)is the core of artificial intelligence and shows excellent superiority to traditional classification methods in finding commonality and distinguishing differences.Nowadays,ML have been widely developed and showed a promising application potential in food analysis field.In this study,Gas Chromatography(GC-MS),liquid chromatography(HPLC)and near-infrared spectroscopy(NIRS)combined with machine learning methods were adopted to analyze Tieguanyin tea,which lead to establishment of quality evaluation models for distinguishing the quality grades of Tieguanyin tea.The main conclusions are as follows:(1)GC-MS was adopted for the qualitative and quantitative analyses of the volatile components in tea samples.Then the cosine angle,K-value clustering,system clustering and machine learning algorithms were used to analyze the volatile components and construct grade model of Tieguanyin tea.The results showed that a total of 40 volatile components were detected in tea samples,and the contents of nerolidol(28.8-35.0%),?-farnesene(23.7-33.2%)and indole(8.4-11.2%)were higher than other ingredients in the tea samples from different production regions with different grade.In addition,there are obvious differences in types and contents of volatile components in tea samples from various production regions.Using the cosine angle method to analyze the similarity of different grades of tea samples,and found that the similarity of the common volatile components in the tea samples decreased as the quality grades decreased.By modeling of volatile components through six machine learning algorithms of Extreme Gradient boosting(XGBoost),K-Nearest Neighbor(KNN),Random Forest(RF),Gradient Boosting Decision Tree(GBDT),Support Vector Machine(SVM)and AdaBoost,the GCMS-XGBoost algorithm model showed the best classification effect as the mean average precision(mAP)reach 84.4%,the coefficient(R~2)is 0.800,and the area under the curve(AUC)is0.782.The accuracy rate of the GCMS-XGBoost algorithm model is significantly higher than that of system cluster analysis(77.2%)and K-value cluster analysis(78.2%),which indicate the proposed GCMS-XGBoost algorithm model has practicable classification application in Tieguanyin tea quality grading.(2)The concentrations of the main ingredients such as epigallocatechin(EGC),epicatechin(EC),caffeine,gallate(GC),Epigallocatechin gallate(EGCG),theanine(L-Theanine),etc.in the methanol extract of Tieguanyin tea were detected by HPLC-UV.And then,principal component analysis and ML were used to analyze and establish the model with the detected characteristic components.The results showed that the contents of EGCG,Caffeine and EGC in all grades of tea samples were significantly higher than that of other ingredients.Furthermore,the contents of EGC,EGCG,caffeine,EC,GCG and ECG in tea samples showed regularity as third grade>first grade>second grade.The matching rates between rating results of principal component analysis method based on HPLC data and professional tea assessors are83.51%-93.23%with mAP of 87.9%,which indicated that the selected chemical components could effectively evaluate the quality grades of Tieguanyin tea.After pre-processed by Norm method,the HPLC data were used to construct quality grade evaluation models by six ML methods.The results showed that the HPLC-XGBoost algorithm model own the best performance with mAP of 98.9%,R~2 of 0.963,AUC of0.905,and excellent ROC and PR curves.The accuracy of the HPLC-XGBoost(98.9%)model is obviously higher than that of classic principal component analysis method,which can accurately evaluate the quality of Tieguanyin tea with unknown grade.(3)The NIRS method was introduced to detect Tieguanyin tea samples,and six preprocessing methods including FD,SD,TD,Norm,Smooth,and SNV were selected to optimized the original NIRS data.Then AutoEcoder,PCA,and HAMNN methods were performed aiming to screen the optimal dimensionality reduction method.Finally six machine learning algorithms such as XGBoost were used to establish the quality evaluation models of Tieguanyin tea.The results showed that the SNV method own the best data preprocessing effect among the six preprocessing methods and the HAMNN dimensionality reduction scheme is superior to AutoEcoder and PCA.The NIRS-XGBoost model shows the best classification performance with mAP of 95.2%,R~2 of 0.901,AUC value of 0.925,and excellent ROC and PR curves.The results indicate that the NIRS-XGBoost model could accurately predict the quality grade of Tieguanyin tea.(4)By comparison of Tieguanyin tea quality evaluation models established by GC-MS,HPLC and NIRS combined with machine learning algorithm,it was found all the models could effectively evaluate the quality grade of Tieguanyin tea,although the evaluation performance of these models have some differences.Among these models,the HPLC-XGBoost model shows the highest mAP(98.9%)and R~2(0.963);the NIRS-XGBoost model own a higher mAP(95.2%)and the highest AUC value(0.925).The comprehensive performance of the HPLC-XGBoost and NIRS-XGBoost models are both more effective than that of models based on GC-MS.It takes about 250 min and 200 min respectively to analyze Tieguanyin tea samples by GC-MS and HPLC methods.Although it takes a long time and consuming some organic solvent,these two methods can obtain information about the corresponding chemical components that determine the quality of tea,which can promote qualitative and quantitative evaluation of tea samples.The quality evaluation system established by the GC-MS and HPLC methods combined with machine learning are instructive for the improvement of tea production technology and meet the needs of scientific research.NIRS method just consumes about 40 min to evaluate the quality grades of Tieguanyin tea,which is obviously shorter than those of methods based on HPLC and GC-MS.NIRS method is a convenient and fast grade evaluation method.However,NIRS method only show spectral data,and cannot directly provide the chemical composition information that related to the quality of Tieguanyin tea.The evaluation method established by the NIRS method combined with machine learning is accurate and convenient,and can meet the needs of rapid evaluation both in the production and market transaction.