| Tea quality was constituted by different sensory quality factors,such as aroma,taste and color.In this study,taking Longjing tea as research object,electronic nose(E-nose),electronic tongue(E-tongue),electronic eye(E-eye)were jointly applied for evaluating the tea quality of different grades,the tea quality of different planting origins and the tea quality of different storage time.The signals from different sources were fused for analysis based on fusion strategies.The evaluating results obtained from independent source and fusion signals were compared,as well as the performance of different strategies were compared either.At the same time,the effect of feature selection methods on fusion strategies were discussed.Eventually,a hybrid feature selection method and corresponding fusion strategy were proposed for synergetically synthesizing the information from E-nose,E-tongue and E-eye.The main conclusions were as follows:(1)E-nose and E-eye were jointly applied to acquire the aroma and image signals of dry tea samples,then the fusion signals of aroma and image were applied for non-destructively detecting tea quality of different grades.Fusion models were set up based on feature-level and decision-level fusion strategies combined with K-Nearest Neighbor(KNN)and Support Vector Machine(SVM)classifiers.It could be found that the KNN and SVM classification accuracy based on independent E-nose and image signals were lower than 90%.The classification accuracy of all the fusion models were higher than 90%.The decision-level fusion combining the SVM results of both E-nose and image signals achieved the highest accuracy of 99%.The results indicated that the fusion of electronic nose and image signal has complementary and enhanced effect on the identification of tea grades.(2)E-nose,E-tongue and E-eye were jointly applied to acquire the aroma,taste and color signals of tea infusion samples,the fusion signals of E-nose,E-tongue and E-eye were applied for synthetically evaluating tea quality of different grades.Feature-level fusion strategy was applied to fuse the information of E-nose,E-tongue and E-eye,and SVM was introduced to construct identification model.It could be found that the identification accuracy based on fusion signals(100%)was higher than that on independent E-nose(93.28%),E-tongue(98.33%)and E-eye signals(90.56%)for identifying tea grades.The results indicated that the fusion of electronic nose,E-tongue and E-eye signals could generate complementary signals which enhanced the identification performance.Meanwhile,the contents of catechins,polyphenols and caffeine in tea were quantitatively predicted based on electronic nose,electronic tongue,electronic eye and fusion signal combined with PLSR(Partial Least Squares Regression)and SVM.The fusion signals achieved better results in predicting the contents of catechins(R~2>0.98),tea polyphenols(R~2>0.97)and caffeine(R~2>0.97)contained in tea than independent signals did.(3)E-nose and E-eye were jointly applied to acquire the aroma and image signals of dry tea samples,then the fusion signals of aroma and image were applied for non-destructively detecting tea quality of different planting origins.By comparing the results provided by fusion signals of E-nose and image with independent ones,it could be found that the fusion signals were not always leading to an improvement for classification result.Pearson correlation analysis(Pearson),information gain(IG)and F-score were introduced for modifying the traditional fusion strategies.Classification models that severally based on feature-level fusion strategies of“feature selection+feature fusion”and“feature fusion+feature selection”and the modified decision-level fusion strategies were constructed.By comparing the performance of the fusion models before and after the modification,and the effect of different strategies,it could be found that the modified fusion models achieved better classification performance than the original fusion models did.The“feature fusion+feature selection”fusion strategy was the most effective one compared with the“feature selection+feature fusion”and modified decision-level fusion strategy.On the other hand,“feature selection+feature fusion”and modified decision-level fusion strategy could prove that the fusion signals contained the information from all the sources.The results suggested that the modification by feature selection methods for fusion strategies could improve the performance of corresponding fusion models,and the proposed fusion strategies have their own advantages.(4)E-nose,E-tongue and E-eye were jointly applied to synthetically evaluate tea quality of different planting origins.Quantum-behaved particle swarm optimization(QPSO)that combined the advantages of“feature selection+feature fusion”and“feature fusion+feature selection”strategies were introduced as modified method for fusing E-nose,E-tongue and E-eye.The performance of Pearson,IG and F-score were compared.It could be found that the classification accuracy of fusion model was improved from 96.25%to 98.75%,98.75%,98.75%and 100%after modifying by Pearson,IG,F-score and QPSO.The QPSO based fusion strategy outperformed Pearson、IG and F-score based fusion strategies.By comparing the prediction results of originals fusion signals and the fusion features optimized by QPSO for catechins,tea polyphenols and caffeine,it could be found that the prediction accuracy was improved from0.9459 to 0.9647 for catechin,from 0.9338 to 0.9715 for polyphenols,from 0.9101 to 0.9459 for caffeine.The results indicated that the invalid information in fusion datasets of E-nose,E-tongue and E-eye could be decreased after feature selection,and further improving the identification and prediction performance of corresponding models.(5)A hybrid feature selection method was put forward which combined the advantages of Pearson、IG、F-score and QPSO.This method and its corresponding fusion strategy was applied to integratedly analyze the 4 datasets provided by E-nose,E-tongue and E-eye for evaluating the tea quality of different storage time.The hybrid feature selection method consists of two steps.Firstly,Pearson analysis,information gain and F-score were combined to preliminarily screen features for a basic subset,and a fine tuning procedure was conducted based on an importance value principle to acquire an efficient subset.Afterwards,QPSO was applied to optimize the fusion efficient feature subsets for an optimal feature subset.The results indicated the optimal feature subset achieved better classification performance than the original fusion signals and the fused efficient subsets from different sources. |
PDF Full Text Request