Raman Spectroscopy Analysis And Application Research Based On Deep Learning

Raman spectroscopy is a kind of scattering spectrum that can reflect molecular vibration information and can provide material molecular "fingerprint" information.As an important analysis technique,Raman spectroscopy has the advantages of nondestructive,fast,and pollution-free,and is widely used in various research fields.In view of the cumbersome process of traditional Raman spectroscopy analysis and the heavy reliance on manual experience for feature selection,the introduction of deep learning methods can optimize the analysis process,improve analysis efficiency,and help the application of Raman spectroscopy to advance toward automation.It greatly improves the accuracy of material analysis and expands the scope of application of this technology in actual production.In order to ensure the stability and reliability of the deep learning model,while avoiding complicated calculation models to reduce analysis efficiency,different deep learning algorithms should be selected for data of different feature types,and the selection of appropriate structures and parameters is critical to the analysis results.This paper uses the excellent learning and computing capabilities of deep learning to study Raman spectroscopy analysis methods based on deep learning algorithms.The main research contents are as follows:(1)For the application of Raman spectroscopy in binary classification tasks,a multi-scale fusion convolutional neural network(GRU-MCNN)classification model is constructed to diagnose and identify the hepatitis B virus(HBV)infection group and healthy control group.It can be used for early screening for HBV infection.The model can be used to fuse multi-scale features and retain time series features,and finally,use the softmax layer to output the classification results.Using real spectral data,experiments were performed on the original data and the preprocessed data.The results show that the GRU-MCNN model has higher accuracy,precision,sensitivity,and specificity on the original data.It shows that the deep learning model can retain more discriminative information while optimizing the analysis process than the explicit preprocessing method.(2)For the application of Raman spectroscopy in multi-classification tasks,and a lightweight deep learning network model is constructed,which increases the width of the network while reducing the depth,which is intended to reduce the computational complexity of the model while ensuring accuracy of classification.In the network structure,a convolutional layer with a convolution kernel of 1*1 is used to reduce the number of input channels,thereby reducing the number of model parameters,and increasing the network width to improve the adaptability to different scales of local spectral features.Using the mineral Raman spectroscopy data set for experiments,the model classification accuracy reached 0.988.According to the comparison of the stability of the classification results,it also reflects that the powerful computing power of the deep learning method enables it to perform better on the Raman spectroscopy data with low consistency and complex classification.(3)For the explanatory problem of the deep learning method,a feature visualization method is used to map the middle layer features extracted by the neural network to the pixel space,which can display the results of the feature extraction of each layer of the deep learning model,and further explain the deep learning model the working mechanism of the internal structure.In the process of identifying key Raman bands related to feature component analysis,the results of feature visualization can play a guiding role.According to the result of feature extraction,it can also guide the model to adjust the structure and parameters based on it,avoiding blind trial and error.Finally,combined with the results of classification experiments prove the reliability of the method.