Research On Raman-NIR Spectra And Multi-spectra Fusion Algorithm For Material Analysis

Raman spectroscopy and near-infrared(NIR)spectroscopy are the most commonly used rapid spectral detection techniques and effective technical means for food component analysis.Raman spectroscopy is a fingerprint spectrum of a substance.There are many advantages,simple,with a high recognition rate and a good model transfer effect for a qualitative analysis.However,Raman intensity is weak,and the signal-to-noise ratio is low.The accuracy and stability of quantitative analysis are not good enough.While NIR spectroscopy is an absorption spectrum with high intensity and good spectral consistency.Using chemometrics modeling,the quantitative detection effect is good,but its model transfer effect is not ideal.In addition,it is a morphological spectrum,lacking of fingerprint lines,and qualitative(material discrimination)analysis is more difficult.In this paper,we studied a multi-spectral data fusion algorithm.The core idea is to establish a Raman-Near infrared hybrid model by merging NIR spectroscopy information into the Raman spectroscopy information.On the one hand,the information of NIR spectroscopy is used to improve the accuracy of the quantitative analysis of Raman spectroscopy.On the other hand,the fingerprint characteristics of Raman spectroscopy are used to improve the transfer ability of the model.Firstly,taking consideration of the signal characteristics of portable spectral detection equipment,we studied some automatic algorithms of spectral data preprocessing,including spectral de-spike algorithm,spectral denoising algorithm,spectral baseline removal algorithm and spectral normalization algorithm.We also wrote C code with the automatic algorithms mentioned above,and successfully transplanted these algorithms to a handmake portable Raman spectrometer.Secondly,we studied two major categories,including machine learning and neural networks,and four types of spectral analysis algorithms,such as decision tree,support vector machine,BP neural network and LSTM neural network.All spectral information is kept to be used in model training process within the establishment of the spectral model as much as possible.The main work is divided into three parts:qualitative analysis,quantitative analysis,and model transfer.The main results are as follows:Qualitative analysis: the binary and multi-class models of Raman spectra are established with decision trees,support vector machine and LSTM neural networks.These models are used to perform qualitative analyses on a single compound.The results show that a correct rate of the binary with decision trees is 98.93,63.28% for multi-class with decision trees.A correct rate of the binary with support vector machine is 100%,91.40% for multi-class with support vector machine.And a correct rate of four-class qualitative analysis using deep learning LSTM neural networks reaches to 100%.Quantitative analysis:(a)a near infrared spectroscopy quantitative analysis model of glucose was established using a decision tree algorithm,with results of R2 = 0.9960,RMSE = 0.0166,MRE = 0.0099.(b)Quantitative analysis models of near infrared spectroscopy,Raman spectroscopy,and Raman-NIR fusion spectroscopy were established,with results of NIR correction set R2=0.9935,RMSE=0.0224,MRE=0.0488,Raman correction set R2 = 0.9246,RMSE = 0.0570,MER = 0.0797,and Raman-NIR fusion correction set R2 = 0.9975,RMSE = 0.0141,MRE = 0.0292.(c)A near infrared quantitative model was established with BP neural network,with the correction set R2 = 0.9962,RMSE = 0.0215,MRE = 0.0439.(d)The quantitative analysis models of near infrared spectroscopy,Raman spectroscopy and Raman-Near infrared fusion spectroscopy were established by LSTM neural network,with results of Raman correction set R2=0.9590,RMSE=0.0058,MRE=0.1505,NIR correction set R2=0.9883,RMSE=0.0298,MRE=0.0586,and Raman-NIR fusion correction set R2=0.9886,RMSE=0.0243,MRE=0.0385.Model transfer:(a)the transfer accuracy rate in the same type of spectral instruments reaches to 100%,using the four-class qualitative analysis model established by the deep learning LSTM neural network algorithm.(b)The quantitative analysis models of NIR spectroscopy and Raman-NIR fusion spectroscopy were established using support vector machine algorithm.The model transfer experiments were carried out in four instruments.The results of model transmitted on the four instruments are 0.7951,0.8664,0.8782 and 0.8819 for NIR model,0.9365,0.9411,0.9303 and 0.9224 for Raman-NIR,respectively.(c)The quantitative analysis models of NIR spectroscopy and Raman-NIR fusion spectroscopy were also established using deep learning LSTM neural network algorithm.The model transfer experiments were carried out in four instruments.The results of model transmitted on four instruments are 0.7768,0.7996,0.8027 and 0.7914 for NIR model,0.5548,0.8444,0.9539 and0.9353 for Raman-NIR,respectively.