Research On Prediction Of Sensory Evaluation Indicators For Unblended Cigarette Based On Data Mining

Tobacco is a consumption product of flue gas, and people can get pleasure by sucking tobacco. Unblended cigarette is an important basis for the formulation of tobacco. As a result, the sensory evaluation of unblended cigarette is the foundation and core of the whole cigarette enterprise product design. However, the process of tobacco burning, pyrolysis and distillation is very short, and smoking experts must capture various characteristics of flue gas in very short time and assess it. It is very complex and difficult to assess the quality of tobacco objectively. Over the years, tobacco enterprises have collected a large number of historical data of sensory evaluation by smoking experts. These data which contain many relationships among unblended cigarette sensory indicators and physical and chemical index are very valuable, but they are stored separately and have poor consistency, low integration degree and low correlation degree. With data mining technology to deal with these data, the great value which they contain and the inherent variation of sensory evaluation can be obtained. Thereby, this approach can provide decision support for the design and production management of the cigarette product, realize digital product design and production of the cigarette and enhance the stability of cigarette production and the market competitiveness. In short, this approach has important practical significance to promote the sustainable development of tobacco enterprises.Based on the historical data of unblended cigarette sensory evaluation, in this thesis, the relationships among 9 sensory evaluation indicators such as flavor, aroma quantity, aroma quality, concentration, irritation, vigour, offensive odor, purity, aftertaste and 8 physical and chemical indicators such as location, grade, total sugar, total nicotine, reducing sugar, total nitrogen, potassium, chloride, are analyzed. In order to reduce the smoking costs and pressure of smoking experts and improve the efficiency and accuracy of the sensory evaluation, the prediction models of unblended cigarette sensory evaluation are established by combining data mining and intelligent optimization methods. The major work of this thesis includes the following three parts:(1) The data of unblended cigarette sensory evaluation are preprocessed, and the scatter plots and correlation coefficient matrix are established with data mining software SPSS. Through the scatter plots and correlation coefficient matrix, it can be found that the different chemical composition has a certain effect on the sensory quality and the correlation between unblended cigarette sensory evaluation indicators and physical and chemical indicators.{2) The fitting and classification prediction models of unblended cigarette sensory evaluation indicators based on BP neural network (BPNN) are established. According to the characteristics of the BPNN, genetic algorithm is used to optimize the weight and threshold of the BPNN, and to obtain more robust models. Through the experiment with these models, it can be found that all models have certain stability, most of the sensory indicators error is within the acceptable range in the fitting models and aroma quantity, aroma quality and irritation have better fitting effect. In the classification models, the classification prediction accuracy of sensory indicators is not high, which may be due to that the sample data is highly nonlinear. Therefore, prediction model based on SVM is established in the fifth chapter.(3) According to the characteristics of unblended cigarette sensory evaluation indicators, SVM models are chosen to forecast the indicators, and the influence of parameters of SVM on the model performance is analyzed. This thesis also used cross validation method to optimize network parameters of the prediction models. Through the experiment with these models, it can be found that the classification prediction accuracy of most sensory indicators is more than 60% and aroma quantity, aroma quality and irritation have better fitting effect. In the fitting models, the effect of BPNN is better than that of SVM. However, in the classification models, the effect of SVM is better than that of BPNN.