Research On Quality Evaluation Of Cigarette Raw Materials Based On Near Infrared Spectroscopy

Entering a new development stage,focusing on promoting the high-quality development of finished cigarettes is the key to enhancing the core competitiveness of tobacco companies.Tobacco is the main raw material for cigarette production and processing,and its quality is related to the quality of factory-made cigarettes.Combined with modern analysis methods,constructing and perfecting a scientific tobacco quality evaluation system is of great significance for realizing tobacco quality monitoring and improving cigarette production efficiency.Based on the near-infrared spectrum data of tobacco,this paper studies the quantitative and qualitative analysis methods of the nearinfrared spectrum of tobacco,as well as the technology related to model transfer,aiming at the problems encountered in the quality evaluation of cigarette raw materials.A tobacco quality evaluation model with high stability and strong generalization ability was established,and the model was shared among multiple instruments.The main research contents are as follows:(1)In order to share the tobacco quality evaluation model on multiple spectral acquisition instruments,improve the applicability of the tobacco near-infrared spectral model.In this paper,aiming at the problem that the function transfer relationship between the master and the slave is difficult to determine due to the non-linear interference caused by the difference between different near-infrared spectrometers and environmental factors,a sparrow search algorithm based on a deep feed-forward network(SSA-DFN)model transfer method for near-infrared spectroscopy.Using a deep feed-forward network to fit the nonlinear function mapping between the spectra collected by different instruments,and using the sparrow search algorithm to initialize the connection weights and thresholds of each layer of the network,not only can quickly obtain the global optimal solution of parameters,Moreover,the convergence speed of the transfer model is improved.To verify the effectiveness of the method,the transfer model of master and slave machine spectra was established by standardizing tobacco spectra,and the spectral effects of SSA-DFN,PDS,and CCA algorithms after transfer were compared.Experiments show that the transferred slave spectrum and the host spectrum are highly consistent,and the prediction results of total sugar and nicotine content are better than other model transfer algorithms,and the model can be shared between multiple instruments,which provides a guarantee for online quality evaluation of tobacco.(2)Spectral data generally have high-dimensional and high-redundancy characteristics,which makes the prediction accuracy of the established spectral quantitative analysis model for the chemical components in tobacco not high,which affects the results of tobacco quality evaluation.Aiming at this problem,a t-distribution stochastic neighbor embedding algorithm(Wt-SNE)based on Wasserstein divergence was proposed,which was used to extract tobacco characteristic information from highdimensional spectral space,and combined with partial least squares method to establish tobacco chemical composition and Quantitative analysis models between spectra.This method uses the Wasserstein divergence to improve the measurement method of the tSNE algorithm for the probability distribution of sample points in two spaces,so that the data after dimensionality reduction retains the characteristic structure of highdimensional spectra,and these characteristic expressions are useful for the quantitative analysis of the chemical components of tobacco has an important role.Compared with the dimension reduction projection of PCA,LPP,and t-SNE algorithms,the results show that the spectral data categories after Wt-SNE dimension reduction are more obvious,and good prediction results have been achieved in the quantitative analysis model of total sugar and nicotine.To a certain extent,the accuracy of the prediction model of tobacco chemical composition was improved.(3)In order to improve the objective authenticity of tobacco quality evaluation and avoid misjudgment of tobacco categories caused by subjective factors during manual sorting and classification,this paper proposes a qualitative analysis method for tobacco based on the improved deep residual shrinkage network(DRSN).To improve the accuracy of qualitative analysis of tobacco near-infrared spectroscopy.This method converts the near-infrared spectrum into a two-dimensional image by using the gramian angular summation fields(GASF),and fully exploits the advantages of the neural network in image processing.The embedded attention module improves the feature attention of the network to the local band of the spectrum.At the same time,two adjustment factors are introduced to optimize the adaptive selection of the noise threshold to reduce the interference of noise in the spectrum.In order to verify the effectiveness and stability of the model,the network model was compared with SVM,RF,CNN,and DRSN,and the results showed that the proposed method achieved high accuracy in the classification tasks of tobacco grade and production area.The rate ensures the accuracy and consistency of tobacco classification and improves the efficiency of online analysis of tobacco quality.