Research On The Discrimination Between Standard Tobacco Spices And Essences And Their Fakes Based On Electronic Nose

Tobacco spices and essences,used as complex and important additives or ingredients in the production of tobacco products,can either impact characteristically distinct aromas and tastes or,to protect the end products from deterioration by microorganisms.Thus,the high quality of spices and essences is the guarantee of eligible end products intended for commerce.However,the quality control of spices and essences remains challenging due to their complex components.Up to now,manual evaluation has been widely adopted to qualify spices and essences.While this method possesses definite subjectivity and is vulnerable to physical condition and subjective will.Another trend in the quality control of tobacco spices and essences is the fingerprint method through modern analytical technology,which can meet the needs of qualitative and quantitative identification.Nevertheless,the utilization of modern analytical technique may be handicapped by the demands for expensive equipment,complicated sample preparation,long test time,and professional operators.In this sense,electronic nose technique,which is low cost,highly sensitive and easy to operate,has been recognized as an efficient method,promising for the introduction into the quality control of tobacco spices and essences.According to the information provided by China tobacco Yunnan industrial Co.,Ltd,FTIR-ATR was used for recognizing 4 groups of samples tested in this thesis.The results showed that there was little difference between the FTIR-ATR spectra of various samples among the same group,with their similarities higher than 98%.However,the tastes of the end-products appeared quite different,which cannot match the demands in practical production.Based on this,we set up an electronic nose system to explore its application in the discrimination between standard tobacco spices and essences and their fakes among four groups.Firstly,WO₃ was chosen as the substrate material and was synthesized by cotton template method,and the surface of WO₃ was modified with Zn,In,Cu and Sn.The results of XRD,SEM and TEM showed that WO₃ synthesized belonged to orthorhombic system with good crystallinity and exhibited highly irregular plate-like,with Zn,In,Cu and Sn uniformly distributed on the surface of WO₃.Subsequently,Radar fingerprints of each sample based on gas sensitive response were plotted with radar mapping tools in Excel.The results showed that the radar fingerprint could only distinguish the standard samples from their fakes among the first group and the fourth group,but failed to realize the discrimination between the standard sample and their fakes among the second group and the third group.Moreover,the principal component analysis（PCA）method was used to process the gas sensitive response in the factor analysis module of SPSS.The results showed that the principal component analysis method can realize the classification of standard tobacco spices and essences and their fakes among four groups.Secondly,three machine learning algorithms including artificial neural network,support vector machine and random forest were used to analyze the recognition accuracy of each group,and the models were further optimized according to the characteristics of each algorithm.For the artificial neural network algorithm,the recognition accuracy of each group was quite different.The recognition accuracies of the first and third groups can reach 100%,while the recognition accuracies of the second and fourth groups were 82.2%and 48.9%respectively.After optimizing the model and getting the best hidden layer node number and the best data dimension of each group,the results of quantitative calculation showed that the improvement of the model by optimizing the hidden layer node number was limited,and the optimization of data dimension can effectively improve the recognition accuracies.The recognition accuracy of the second group was improved from 82.2%to 84.6%,and the recognition accuracy of the fourth group was improved from 48.9%to 91%.As to support vector machine algorithm,the recognition accuracies of the four groups were more than 92%,and the recognition accuracies of the first group and the third group were 100%.After optimizing the model,the best penalty parameters and the best data dimension of each group were obtained.The results of quantitative calculation showed that the optimizations can slightly improve the recognition accuracy of the second group,but greatly reduced the recognition accuracy of the fourth group.In terms of random forest algorithm,the recognition accuracies of samples among four groups were more than 92%,and the recognition accuracies of the first and fourth groups of samples were 100%.After optimizing the model and getting the optimal number of decision trees for each group,the results of quantitative calculation showed that the improvement of recognition accuracies was limited.By combining electronic nose technology with principal component analysis and machine learning algorithm,we can distinguish standard tobacco spices and essences from their fakes among four groups,which can be used for references by the research of the quality control of tobacco spices and essences.