Data Fusion Of Near Infrared Spectroscopy And Electronic Nose And Its Application In Tobaccos

To increase the competitiveness of tobacco enterprises in tobacco market at home and abroad,the Near Infrared Spectroscopy(NIR)with abundant information on structures and compositions,and Electronic Nose(EN)with signitures of biological smell,have been employed into the instant detection for tobacco ingredients,in the purpose of online supervisory control for tobacco quality as well as precise production and hence reducing production cost.However,it's difficult to identify the chemical compositions of tobacco,since tobacco owns a complicated chemical system.It's not sufficient to solve practical problems of tobacco by single dimension data sourced from NIR or EN,such as discrimination of tobacco parts and quantitative predictions for chemical compositions of low contents.For the deeper researches of complex chemical system of tobacco,the data of two dimensions from NIR and EN have been fused to conduct the studies of the identification for the tobacco chemical system and ingredients.The main researches and results are listed below:1.Based on the data of NIR,discriminant model of classifications for tobacco parts has been conducted through Support Vector Classification(SVC).During the research,the influences of spectrum areas,dimensionality reduction methods,feature selections,modelling methods,selection of Kernel functions have been discussed detailly.For NIR-SVC model,the accuracies of training set,leaving-one-out cross-over,prediction of testing samples are 100.0%,100.0%and 63.33%,respectively.The result shows that the prediction accuracies are not satisfactory only through the NIR data.2.Based on the data of EN,discriminant model of classifications for tobacco parts has been conducted through Support Vector Classification.During the research,the influences of sensory times,different sensors,dimensionality reduction methods,feature selections,modelling methods,selection of Kernel functions have been discussed detailly.For EN-SVC model,the accuracies of training set,leaving-one-out cross-over,prediction of testing samples are 100.0%,99.22%and 46.67%,respectively.The result shows that the prediction accuracies are not satisfactory only through the EN data.3.To increase the prediction accuracies of testing samples,the data of two dimensions sourced from NIR and EN have been fused to conduct the NIR-EN-SVC model in the extent of multiple dimensions.The accuracies of training set,leaving-one-out cross-over and prediction of testing samples are 95.31%,79.61%,81.67%,respectively.The result shows that the information contained in the fused model is obviously much more affluent than the single ones.The prediction accuracies using the fused data are satisfactory.4.The fused data also have been applied to the discrimination for the vintages of tobacco,and the quantitative predictions for the potassium content.For the model of the discrimination for the vintages of tobacco,the accuracies of training set,leaving-one-out cross-over and prediction of testing samples are 100.0%,98.50%and 90.00%,respectively.For the model of the quantitative predictions for the potassium content,the average relative errors of training set,leaving-one-out cross-over and prediction of testing samples are 9.11%,10.45%,10.55%,respectively.The performance of model using the fused data exceeds those using the NIR or EN data.In short,in the consideration of the chemical systems of tobacco,the models with better performances could be established due to the data fusion from multiple dimensions and could be applied into the discriminations for tobacco parts and vintages as well as the quantitative predictions for the potassium content.We hope that this study could provide reference contents for the descendent researches and new tools for the instance detections of relevant properties.