Research On Data-driven Methods For Cigarette Formulation Maintenance

Design of cigarette formula is the core and key task in the design and production process of cigarette products.A cigarette formula is made up of a variety of different single tobaccos with appropriate ratios.Adequate stock of single tobaccos plays an important role in the stable production of cigarette products.However,the quality of cigarette products may be fluctuated with changes of the single-tobacco composition of cigarette formula for a period of time by factors such as market cost,inventory composition,and purchase time.In order to maintain the stability of the quality of the cigarette products,it is necessary to replace single tobaccos and adjust the proportion of single tobaccos in cigarette formula.However,it's very complicated and difficult to maintain cigarette formula.In traditional cigarette formula maintenance process,most tobacco companies mainly rely on sensory evaluation by experts based on empirical analysis or repeated experiments.But it is time-consuming to maintain cigarette formula only by human sensory evaluation which will lead to low production efficiency and increased production costs.Therefore,how to intelligently maintain cigarette formula has become an important issue in tobacco companies.With the development of computer and artificial intelligence technology,people began to enter the era of information.Tobacco companies also have accumulated a huge amount of formula data.Through data mining and analysis,valuable knowledge can be extracted from these data for the maintenance of cigarette formulas.This thesis established three maintenance models of cigarette formula based on Hybrid Collaborative Filtering(HCF),Non-negative Matrix Factorization(NMF)and Two-step Kernel Regular Least Squares(TKRLS).The implicit compatibility rules between single tobaccos in formula data can be excavated by using these methods.And then the optimal replacement of single tobacco can be obtained according to the similarity-based heuristic rules.The main contents of this thesis are given as follows:(1)Firstly,the basic principles of user-based and item-based collaborative filtering methods and similarity calculation methods are analyzed in detail.The formula similarity matrix and single-tobacco similarity matrix are constructed.With the advantages of filtering the useless single-tobacco collocation and effectively using the information of similar formula and single tobaccos,this paper established a cigarette formula maintenance model based on HCF.(2)Firstly,the basic principles,characteristics and application scenarios of NMF are analyzed.Because NMF can deal with the matrix sparse problem of collaborative filtering,and has an intuitive semantic interpretation,a linear model of single-tobacco prediction is established based on the combination of formula data and NMF,which can calculate the model matrix.After that,the model of cigarette formula maintenance based on NMF is established.Finally,the experimental verification of NMF is carried out and compared with the HCF method.(3)A two-step kernel least squares method is established based on the knowledge and characteristics of least squares regression,kernel method,regularization,migration learning and two-step method.Then,based on the single-tobacco prediction model and the application of information of the cigarette formula and chemical composition of the single tobaccos,the cigarette formula maintenance model is established based on TKRLS.Finally,TKRLS is compared with NMF and analyzed by the experiment.(4)The recommended single-tobacco item list can be obtained by using the above three methods,and then two alternatives are proposed,namely one-to-one replacement and many-to-many replacement.Three different many-to-many replacement methods are also proposed for different replacement situations to select the final replacement of single tobacco.