| In order to improve the market competitiveness of cigarette products and expand the attention of brands,the quality stability and monitoring of cigarette products are very important.However,for a long time,the intrinsic quality characterization of cigarette products mainly through sensory evaluation,the accuracy and stability of the results are affected by the individual differences of the users.It is difficult to realize online monitoring because of the lack of quantitative description method of intuitive image.In addition,the leaf blend formula is the key to ensure the quality of the product,and the stability of the formula is conducive to the stability of cigarette product quality.However,due to the changes of the international and domestic market prices and the objective factors such as year,climate and soil,some single tobacco in the cigarette formula tobacco is short.It is one of the key links of cigarette product maintenance to select the appropriate stock tobacco for replacement.At present,the traditional methods of tobacco quality evaluation mainly use the physical and chemical index of tobacco leaves,qualitative description of the appearance quality of tobacco leaves and sensory evaluation of the experts.However,the evaluation of tobacco quality by scoring still has some problems such as lagging,incomplete,objective,heavy workload,time and raw materials loss,which can not meet the practical problems such as rapid production of enterprises.Therefore,how to evaluate the stability of cigarette products objectively and efficiently and automatically recommend alternative tobacco from the existing tobacco in stock is an important part in highlighting the industry level and realizing process control.In this paper,combined with near infrared spectroscopy and chemometrics methods,based on the experimental materials provided by China Tobacco Guizhou Industrial Co.,Ltd,firstly,pattern recognition and similarity measurement of finished cigarettes are performed from the perspective of variable dimensionality reduction.Secondly,the analytical design of leaf group formula was carried out,and the analytical model of leaf group formula was established.Then,combining with the information of stored tobacco leaves,the similarity of leaf group formula tobacco leaves was measured,and the replacement design was carried out.At last,Matlab GUI is used to build two assistant modules,which are leaf blend formula analysis and leaf blend formula design.The main results are as follows:1.Since the SIMCA and PLS-DA models in the national standard can not realize the qualitative identification of finished cigarettes,three supervised pattern recognition methods are designed to evaluate the quality stability of cigarette products from the perspective of variable extraction.From the division of sample set,the determination of spectral preprocessing method,the extraction of spectral data variables and the determination of pattern recognition method,the analysis model of finished cigarette is established,and the best discriminant model is determined by the final discriminant accuracy of the model.The results show that in the first method,CWT is used to preprocess NIR spectrum,PPCA is used to reduce data dimension and linear kernel function is used to establish SVM method which the correct classification rate of the prediction model can reach 97.20%.The second model is based on the qualitative analysis variable selection reported in the literature,the accuracy of PPCA-LDA and SVM correction model based on WTSD and baseline correction can reach more than 99%.The third model based on the new variable extraction method proposed by the analysis of variance,after optimizing the key parameters of the algorithm,the correct classification rate of SVM and LDA analysis model can be improved by 15% compared with the model built by the analysis of variance.The results show that the identification accuracy of the best model based on the three variable extraction methods is higher than the analysis methods reported in the literature,which can be used to identify the finished cigarettes and provide objective guidance for the evaluation of the quality stability of cigarettes.2.The supervised pattern recognition method can analyze the fluctuation of cigarette brands among different batches and monitor the product quality,but it can not directly reflect the similarity between brands.The similarity measurement models of SAM,SCM,MDM,EDM,Hsim and Close are established based on the nonlinear dimension reduction method.The similarity measurement of cigarette based on LLE dimension reduction algorithm using SAM similarity measurement method can reach85%.Compared with the supervised pattern method,this technology is simpler and can be applied to the similarity measurement of large quantities of samples.A joint evaluation method based on five similarity algorithms is proposed.The similarity between brands by calculating the size of similarity and intimating relationship of cigarettes through sorting the similar finished cigarettes,it has practical guiding significance for the development of new products and the optimization and quality improvement of old products.The Fisher quality projection method is used to realize the visualization of similarity measurement through the two-dimensional and three-dimensional graph,and the similarity of each cigarette product can be intuitively seen from the space.It can provide scientific and intuitive basis for the relative positioning of cigarettes in the market and formula aided design.3.Three nonlinear dimensionality reduction algorithms and three feature selection methods are used to measure the similarity of reconstructed leaf blend.After the spectrum pre-processing,combined with different variable extraction methods,six similarity evaluation algorithms were used to measure the similarity between the reconstructed leaf blend and the pure leaf blend.Among them,the SAM similarity model based on LLE dimension reduction combined with CWT spectral pretreatment can achieve 80% of the measurement value of reconstructed leaf blend,which can provide guidance for leaf blend formulation design.4.Based on the formula regression and nonlinear multivariate optimization algorithm,after determining the corresponding objective function and constraints,the fmincon function in Matlab is used to analyze the leaf blend formula.According to the idea of leaf blend formula analysis,the process of iterative optimization has been continuously carried out,and 8 leaf group formulas out of 11 leaf groups have been successfully analyzed.Based on the similarity measurement method,the similarity study was conducted from two aspects of tobacco leaf and leaf blend.The three tobacco leaf raw materials of G brand leaf group were replaced by similarity.After the arrangement and combination of similar tobacco leaves,the absolute deviation of the conventional index between the selected replacement leaf blend and the original leaf blend was no more than 0.3 through module similarity measurement and reconstruction leaf blend similarity measurement.This method can be used to guide formula maintenance.5.In order to realize the human-computer interaction easily and realize the calculation and result display function quickly,the graphical user interface(GUI)of Matlab is used to develop and design the leaf blend formula design model.At present,the two module platforms of leaf blend formula analysis and maintenance have been successfully built. |
PDF Full Text Request