Study On Qualitative And Quantitative Analysis Method Of Raman Spectroscopy Based On Deep Learning

Raman spectroscopy has many advantages,such as simple sampling,rapid detection,non-destructive to samples,wide range of applications and so on.Qualitative identification and quantitative analysis of substances is an important task of Raman spectroscopy.The traditional Raman spectral analysis method should first preprocess the Raman spectrum,then extract the specific peak information or use principal component analysis to extract features.Manual intervention in the process of spectral signal processing will inevitably lead to data loss or change,which will affect the accuracy of the analysis results.As an important deep learning algorithm,convolutional neural network has the characteristics of local connection and weight sharing,which enables it to accurately and efficiently analyze multi-dimensional and large amounts of data.In this paper,the qualitative and quantitative models of Raman spectroscopy are established by using convolution neural network.the main research contents are as follows:Raman spectra of three estrogen powders(estrone,estradiol and estriol)were collected for qualitative identification.Surface-enhanced Raman spectroscopy of melamine with different concentration gradients were collected for quantitative modeling.Three kinds of Raman spectral data augmentation methods are designed to expand the spectral data and construct spectral data set,which is convenient to train the neural network model and improve the generalization performance of the model.Based on LeNet-5 classical neural network model,a one-dimensional convolution neural network classification model for Raman spectral signals is proposed to identify the Raman spectra of three kinds of estrogen.Through a large number of simulation experiments,the hyper-parameters and training process of the proposed classification model were optimized and the training was completed.The performance of proposed model is evaluated by comparison with various traditional Raman spectral classification algorithms.The results show that the one-dimensional convolutional neural network model proposed in this paper can quickly and accurately classify Raman spectral data without preprocessing and feature extraction steps,which can simplify the spectral analysis process.Compared with the traditional Raman spectral classification method,the convolutional neural network model is less affected by the noise of spectral measurement and has stronger robustness,which is suitable for the analysis of Raman spectral signals with strong noise measured in complex situations.On the basis of spectral classification model,the structural were adjusted and the quantitative prediction model of Raman spectra was built by convolutional neural network.The loss function of the model was improved,the training parameters of neural network model were optimized and the training was completed,and the Raman spectral concentration of melamine was predicted quantitatively.Compared with the traditional Raman spectral quantitative modeling method,the convolutional neural network model has a simple analysis process,better fitting effect and accurate prediction results.Experiments show that convolution neural network is suitable for the analysis of Raman spectral data,and the process is simple and the results are accurate.The convolution neural network model proposed in this paper provides a new perspective for qualitative and quantitative analysis of Raman spectroscopy.