Research On Material Identification Method Based On Raman Spectroscopy In Combination With Deep Learning

Raman spectroscopy has been widely used as a material analysis tool and has formed an extensive spectral library.However,the spectral collection process may be affected by various interference factors,resulting in a decrease in the quality of Raman spectroscopy.In order to solve this problem,traditional Raman spectral analysis methods require preprocessing of the spectra,including baseline correction and noise removal.However,these methods require significant manual operations,leading to low efficiency.Therefore,this thesis proposes a Raman spectroscopy combined with deep learning method for material identification,aiming to improve efficiency and leverage the big data advantages of spectral libraries.The main research contents are as follows:(1)A simple data augmentation deep learning method is proposed for the Raman spectroscopy identification problem with complex baselines,aiming to achieve Raman spectroscopy identification without the need for baseline correction.A data augmentation method was employed,where simple functions were added to the spectra in the database to generate the training set.The trained model achieved 100% recognition of 20 different minerals with complex baselines,without the need for additional baseline correction.The influence of training set size on model performance was examined,revealing that a data size exceeding 300 samples resulted in a testing accuracy of 100%.Additionally,the impact of linear and sinusoidal function data augmentation methods on recognition performance was explored,and the results demonstrated the model’s stability in handling variations in the training set.(2)Aiming at the problem of Raman spectral noise reduction with high intensity noise,the noise reduction task of low signal-to-noise ratio spectrum is solved based on the deep learning model.The training set was obtained by using a data augmentation method that overlaid Gaussian white noise on normalized ideal spectra.The trained model successfully denoised Raman spectroscopy of 20 types of minerals,with a signal-to-noise ratio(SNR)improvement of over 21 times compared to the SNR before denoising.In addition,it was found that data normalization significantly affected the performance of the model and could improve the training effect.The study also investigated the impact of the size of the training set and the data augmentation method on the performance of the model and found that the optimal denoising performance was achieved with a training set of 2000 spectra using the data augmentation method that normalized the spectra and added Gaussian white noise to achieve an SNR of 30.(3)A research study was conducted to address the recognition of Raman spectra with complex baselines and high-intensity noise.The study focused on investigating the training methods for a deep learning joint model.By employing a data augmentation method that involved adding Gaussian white noise to spectra with linear function baselines,the study achieved 100% accurate recognition of Raman spectra from 20 different minerals that contained complex baselines and high-intensity noise.Notably,the recognition was achieved without the need for manual preprocessing.Furthermore,the study discovered that incorporating a data recovery module enhanced the joint model’s ability to recognize Raman spectra with low signalto-noise ratios.Lastly,the joint identification model was evaluated using a completely independent test set,which resulted in a testing accuracy of 100%.These findings indicate the successful implementation of a deep learning joint model for Raman spectroscopy identification without the requirement for manual preprocessing steps.This thesis demonstrates the application and advantages of deep learning in Raman spectroscopy analysis.The proposed Raman spectroscopy joint deep learning identification method can achieve intelligent and fast Raman spectroscopy identification without relying on manual operations,thereby improving identification efficiency.This method has significant implications in fields such as materials and chemistry,providing new ideas for related research.