Research On Deep Learning Algorithm For Near Infrared Spectral Feature Extraction Of Tobacco Leaf Materials

Because the Near Infrared Spectroscopy(NIRS)of tobacco leaf material belong to high-dimensional data,it has the characteristics of overlapping and non-linearity and so on.It is easy to appear dimension of disaster and can not clearly show the sample of the active ingredient information.In the qualitative analysis of the spectrum,if we use the full spectrum to model,it will inevitably bring some noise and interference to analyze the spectrum.Therefore,in order to solve these problems,this paper presents an effective method of filtering and dimensionality reduction of the original spectrum.In this paper,it center on the analysis of near infrared spectroscopy of tobacco leaf,the main contents are as follows:1.In this paper,at the beginning of the characteristics of Near Infrared Spectroscopy of tobacco leaves were analyzed,such as spectral processing,optimal wavelength variables and feature extraction were discussed.The basic idea and realization process of the depth learning algorithm were introduced.A targeted analysis of autoencoder,sparse autoencoder and convolution neural network structure design and algorithm derivation has provided a theoretical support for the follow study.2.Through the analysis of the Near Infrared Spectroscopy of tobacco raw materials,we can see that if the thousands of characteristic variables of the whole spectrum are modeled,the calculation is large.In this paper,we use the competitive adaptive weighting algorithm(CARS)to extract the characteristic wavelength of the near infrared spectrum of tobacco leaf material,combining with the partial least squares algorithm,and establish a correction model of nicotine.Through comparison and without band selection and with no information variable elimination method for band selection to establish a model.The study showes that the model established by CARS method can reflect all the information of 256 characteristic variables in the whole band with 32 wavelength points atthe same time,only three main components are extracted,and 12 in the whole band,which greatly reduces the model in contrast,the CARS method established a strong external verification capability,with a small relative average error,which reduced unnecessary noise,and multiple collinearity between the various wavelength points,increasing the model established.The model is better and the robustness is better after the combination with PLS,and the feasibility and anti-jamming ability of CARS method in the optimization of the characteristic wavelength of near infrared spectrum of tobacco leaf material are verified.3.In this paper,we propose a method of extracting SAECNN from sparse automatic coding network and convolution neural network by using the depth learning algorithm to extract the Near Infrared Spectroscopy of tobacco leaf raw material,which is different from the previous feature extraction method.Through the SAE pre-training network,the CNN structure and internal parameters are optimized,and the visual comparison between PCA algorithm and ISOMAP algorithm is analyzed.The result showed that the highest accuracy rate and average correct rate of SAECNN algorithm were 94.47% and 94.22%,which is higher than 94.28% and 94.04% of the ISOMAP algorithm and 95.47%and 95.13% of the PCA algorithm respectively.It proves that the SAECNN method is feasible and effective for the Near Infrared Spectroscopy of tobacco leaf raw material,and it also provides a new analysis way for Near Infrared Spectroscopy.4.Finally,it is necessary to summarize the work of this essay,and point out deficiencies and areas to be improved,and learn to explore the direction of the next step.