Research On Spectral Classification Based On Deep Learning

The spectrum carries a large amount of molecular fine structure and characteristic information.Each biomolecule or sample will carry unique spectral characteristic information,which is essential for disease screening and qualitative analysis.How to carry out feature extraction at different levels of the spectrum and complete spectral classification and recognition stably and efficiently has become a major challenge for further analysis and research of the spectrum.At present,deep learning technology has been successfully applied in the fields of biomedical auxiliary diagnosis,text classification and chemometric analysis,which provides solutions for the accurate identification and classification of spectra.The traditional machine learning classification algorithm has limited learning ability,and feature extraction of manual intervention and excessive preprocessing operations can easily misses part of the detailed information of the spectrum,which affects the stability and reliability of the algorithm.Therefore,based on deep learning algorithms and various types of spectral data as the research object,this paper establishes a hepatitis B screening model based on long and short-term memory neural networks and a material spectrum classification model based on convolutional neural networks.The specific summary is as follows:(1)Taking human serum Raman spectra as the starting point of the research,collecting serum Raman spectra of hepatitis B patients and non-hepatitis B patients,a classification model based on deep learning is proposed,which completes the classification and identification of hepatitis B.The collected serum Raman spectra were preprocessed,and principal component analysis was used to retain 99.99% of the original data to extract the main features.Established a hepatitis B screening model of long and short-term memory neural network,and compared the model with recurrent neural network,convolutional neural network,multi-layer perceptron,support vector machine,random forest and linear discriminant analysis.Achieved a classification accuracy of97.32%.The experimental results show that serum Raman spectroscopy combined with deep learning classification algorithm has certain feasibility in hepatitis B screening.(2)Aiming at improper preprocessing operations that will change the characteristic information carried by the original spectra,a spectral classification model that simplifies the data preprocessing steps is proposed.The model can start from the original spectral signal,and the convolutional neural network used integrates spectral preprocessing,feature extraction,and classification modeling into a whole.Through training the model,the impact of improper preprocessing operations on the spectral classification model can be reduced.In this paper,five publicly available material spectrum data sets are used to compare the effects of various preprocessing strategies on the classification performance of the model.The experimental results show that on the five public spectral data sets,the spectral classification model can retain the characteristic information of the spectral data and further optimize the preprocessing steps of the spectrum.(3)In order to fully extract the characteristic information of the spectrum from different levels,this paper designs a network structure combining the Inception module and the residual block based on a one-dimensional convolutional neural network.The Inception module increases the width of the network and combines the two convolutional channels after processing.As a result,the characteristic information of the shallow and deep layers of the spectrum can be better extracted,and the residual block speeds up the convergence speed of the network model by means of identity mapping,and realizes the superposition of information.Finally,use the Soft Max classification function to complete the classification and identification of the material spectrum.