Prediction Of Cigarette Smoke Index Based On Support Vector Machine

The formula of tobacco about cigarette products must be adjusted in order to maintain the stability of product style, cost, and indices of smoke. However, the indices of the changed formula won’t be known until the process of production is completed and the indices have been tested. However, it is too late because when the content of tar and other chemical components is overproof, the products will be obsoleted for failing to meet national standards, which may cause serious economic losses to the enterprise. Therefore cigarette companies need to predict the content of the smoke index of finished cigarettes before the process of production begins.Traditional forecasting methods based on the precise mathematical model are difficult to provide a basis for the production because of the nonlinear, high-dimensional and random characteristics of the problem of smoke index prediction. Therefore, support vector machine(SVM) method is used in this thesis. SVM has the incomparable superiority compared with other prediction methods in solving nonlinear, high-dimensional practical problem with small sample size. The main research work in this thesis is as follows:First, the important elements for a given smoke index are listed. the advantages, disadvantages and the suitable problems of different forecast methods are compared. The research status of smoke index forecasting are summarizes. The input variables are chosen for SVM according to the correlation analysis of chemical and physical indices of tobacco, cigarette auxiliary materials and smoke indices. On this basis, SVM is put forward to forecast the smoke index.The smoke indices of unblended cigarette are forecasted based on the correlation analysis. The training set and test set are formed and forecast models are set up by organizing the unblended cigarette data from cigarette enterprise. On this basis, the parameters for the prediction model of smoke index are optimized with the grid method and the prediction results are compared with BP nepal network. Experimental results show that the prediction of SVM with grid method in optimizing the parameters is fine. Error indicators meet requirement and are better than that of BP nepal network. The MAE of tar, smoke nicotine and CO are respectively 0.175%,0.3% and 1.71%.Difference between unblended cigarette and finished cigarette in tobacco and cigarette auxiliary materials makes it more difficult to forecast the smoke index of finished cigarette. The parameters are optimized for SVM by Genetic Algorithm (GA) based on the prediction model of smoke index. Cross operator and mutation operator are designed. Suitable crossing-over rate and aberration rate are obtained by experiments. The smoke indices are forecasted with the optimized parameters and the prediction are compared with grid method. Experimental results show that the prediction of SVM with GA is better than that of SVM with grid method. The MAE of tar, smoke nicotine and CO are respectively 3.09%,4.51%,3.85%. SVM with GA not only has accuracy and stability, but also has generalization performance.At last, the idea of filtration efficiency is put forward aiming at the smoke index of finished cigarette. Compared with conventional methods, the idea of filtration efficiency can improve prediction effect of tar in some degree. Filtration efficiency broadens research field and offers some help for design and manufacturing in cigarette enterprises.