Research On Raman Spectral Modeling And Model Transfer Based On Deep Learning

Raman spectroscopy is widely used in agriculture,food,medicine,chemical industry and other fields because of its rapid,non-destructive and high sensitivity.However,in the practical application of Raman spectrum detection,it is often encountered that the model established on one instrument can not be used on another instrument,which greatly restricts the development and popularization of Raman spectrum detection technology.A common strategy to solve the above problems is model transfer.However,most of the model transfer methods require the preparation and preservation of standard samples,which is complicated and tedious.The deep learning based Raman spectroscopy modeling and model transfer method proposed in this paper solves this problem well.Specifically,the research content of this paper is divided into the following parts:(1)Raman spectrum modeling method based on convolutional neural network: In order to solve the problem that the traditional linear multivariate correction algorithm has insufficient prediction ability in nonlinear system,this paper proposes a Raman spectrum modeling method based on convolutional neural network.Compared with the traditional partial least squares algorithm and support vector regression algorithm,the proposed method achieves better model prediction performance and model robustness by virtue of the good nonlinear fitting ability of convolutional neural network.At the same time,this paper also explores and verifies the extraction mechanism of Raman spectrum characteristic peaks of convolutional neural network by using the Class Activation Map method,and obtains certain interpretability.(2)Raman spectral model transfer method based on convolutional neural network:On the basis of Raman spectrum modeling based on convolutional neural network,a new method of Raman spectrum model transfer based on convolutional neural network is proposed.This method makes full use of the abundant prior information of the master instrument,which effectively solves the problem of insufficient label samples from the slave instrument,and achieves good model transfer effect.(3)Raman spectral model transfer method based on wavelet transform for double branch network: Considering the multi-scale characteristics of Raman spectrum and the characteristics of time-frequency cooperative analysis and multi-scale decomposition of wavelet transform,the time-frequency graph of Raman spectrum after wavelet transform is proposed as another modeling data.In addition,in order to extract the characteristic information of the components to be measured better,this paper proposes a dualbranch network model,in which a Raman spectral sequence and the corresponding Time-Frequency graph of Raman spectrum are used as input data for model training.Through experimental analysis and verification,the modeling effect and model transfer ability of dual-branch network model are further improved compared with single-input convolutional neural network.The main innovation of this paper lies in the application of convolutional neural network and double branch network models in deep learning to the modeling and model transfer of Raman spectral data.On the one hand,the prediction accuracy and robustness of the model are improved in Raman spectrum modeling;on the other hand,the preparation of standard samples is avoided in Raman spectrum model transfer,and only a small number of samples with labels of the slave instrument can achieve good model transfer effect.This paper provides a new idea for Raman spectroscopy modeling and model transfer.