Study On Digital Evaluation Method Of Cigarette Raw Materials Based On Near Infrared Spectroscopy

The quality stability of cigarette products is the core of the sustainable development of tobacco enterprises,which often depends on the quality of its raw tobacco leaves.The quality assurance of tobacco leaves is also the key to improve the service quality of related enterprises.Traditional tobacco quality evaluation mostly uses chemical methods to measure the internal chemical components of tobacco leaves and manually evaluate smoking,which is subjective and ineffective,and consumes a large number of raw materials,seriously affecting the long-term development of enterprises.Therefore,tobacco enterprises urgently need to combine intelligent methods to evaluate the quality of tobacco leaves efficiently and objectively,so as to ensure the stability of the whole chain of product quality in the manufacturing process.Aiming at the problems encountered in the quality inspection of tobacco enterprises,this paper studies a fast and reliable digital evaluation method system based on the near infrared spectrum data of tobacco leaves,its internal chemical composition,grade and other data,combined with spectral quantitative analysis and qualitative analysis techniques and deep learning methods.The specific research contents are as follows:（1）Aiming at the influence of high dimension,nonlinearity and a large amount of noise on quantitative modeling of near-infrared spectrum,this paper introduces the depth auto-enconder network into spectral feature learning,proposes a feature extraction method 1D-BCAE based on improved convolution auto-enconder network,and applies it to quantitative modeling of near-infrared spectrum of key indexes of tobacco leaves to solve the problem of learning nonlinear relationship among high-dimensional features.In this method,firstly,one-dimensional convolution kernel and pooling window suitable for spectral data are used to extract features.Secondly,Basic Block module is added in the coding process,which reduces the number of parameters and the amount of calculation,and at the same time,reduces the influence of noise and nonlinear features in the spectrum.By designing a corresponding connection structure,the corresponding encoder and decoder are connected,and the parameters of each module in the encoder are transferred to the corresponding decoder,which reduces the loss of detail features in the process of network training.The effectiveness of this method is verified by comparing the reconstruction error and root mean square error in experiments.The quantitative models of nicotine and total sugar in tobacco leaves are established by using the features extracted by full spectrum segment and principal component analysis（PCA）,convolutional auto-enconder network（CAE）and 1D-BCAE combined with partial least squares（PLS）respectively.The results show that 1D-BCAE algorithm can effectively learn the internal structure and nonlinear relationship in high-dimensional data,and the established model has better performance.This method completes the effective extraction of spectral information of components to be tested,and realizes the accurate prediction and output of intrinsic chemical components of tobacco leaves.（2）In this paper,based on one-dimensional convolution,a tobacco grade recognition method based on Ghost lightweight convolution neural network is proposed.The VGG16 network is improved,and the square matrix convolution kernel and pool window are changed into vector convolution kernel and pool window suitable for one-dimensional spectral data.The network structure is simplified,and a new Ghost module is proposed to replace the multi-layer convolution superposition structure,and a Ghost-VGG16 model is obtained,which can extract spectral data deeper and prevent the gradient from disappearing.Through batch normalization,the problem of network efficiency reduction caused by the dispersion of internal data distribution after convolution calculation is prevented.Taking the collected near infrared spectrum data of 650 tobacco samples of 5different grades from different producing areas as the data set and the improved neural network as the model,the classification accuracy of the training set and the testing set are 98.0% and 97.3% respectively.Experiments show that the improved model can effectively learn the distribution structure of the original data,and accurately classify the near infrared spectrum data.Using the improved model to classify tobacco grades can get better accuracy.Compared with other classification models,the classification accuracy is better,and the classification efficiency of the model is high.To some extent,it solves the error caused by manual classification in tobacco producing areas,reduces manpower output and improves efficiency.