Prediction Methods Of Cigarette Sensory Index Evaluation Based On Classification Algorithms

With the impact of economic globalization,the techniqual request of cigarette product development in tobacco industry of our country becomes higher and higher.In the manufacture process of a tobacco company,the cigarette formulation that can meet the request of product sensory quality needs to be designed based on unblended cigarette with different chemical components.Therefore,it is important for cigarette manufacturing to study the correlations between cigarette chemical components and sensory quality.Currently,tobacco companies evaluate cigarette sensory quality by the smoking of cigarette experts,which is obviously inefficient and unstable.Based on a large number of data accumulated in practical production,tobacco companies urgently need data mining methods to make intelligent evaluation for cigarette sensory quality.There have been some experts in relevant fields getting study results about intelligent evaluation of cigarette sensory quality using single data mining algorithms such as BP neural network or SVM and rough sets.However,the existing methods rarely use classification methods to make prediction.Moreover,there is no systemic comparision,analysis and ensemble learning of multiple classification algorithms.A software system is also needed,which can combine theory study and guide enterprise application.For these reasons,with the practical sensory evaluation data of cigarette product,this thesis studies the prediction methods of cigarette sensory index evaluation based on classification algorithms.The major work includes the following three parts:(1)Data preprocessing is made according to the characteristics of the practical sensory evaluation data of cigarette product,including data discretilizaion and normalization.Prediction models for sensory index of cigarette product are built based on six different classification algorithms,including decision tree ID3,decision tree C4.5,k-Nearest Neighbor,BP neural network,Support Vector Machine and Naive Bayes.Systemic comparisoin and analysis are made about the performance of these six algorithms when used in the process of cigarette product sensory index prediction by a large number of experiments,which show that SVM can get the best effect.(2)Based on the former studied prediction models of classification algorithms,prediction models for sensory index of cigarette product are built using classical ensemble learning methods Bagging and Boosting.With the characteristic of data sample,Multiple Classifier Systems(MCS)is designed to make classification prediction.The classification prediction performances between single classification methods and ensemble learning methods,and among different ensemble learning methods are compared by a lot of experiments,which show that for some algorithms,Bagging can get obvious accuracy enhancing;Boosting has no effect and even lead to lower accuracy and MCS can get better effect than that of any single classifier which is used to bulid the ensemble system.(3)Based on the classification algorithms studied in this thesis and the existing results of project group,an auxiliary decision-support system of cigarette optimization using Matlab GUI is designed.Five modules are built for experiments and prediction,including unblended cigarette physical and chemical index to cigarette gas index,unblended cigarette physical and chemical index to sensory index,cigarette product physical and chemical index to cigarette gas index and cigarette product physical and chemical index to sensory index.Moreover,an example is applied using practical data to prove that the relative function request has been met.