DCGAN-based Raman Spectral Classification Method And Applied Research

Drug testing technology can be effective in preventing the harm to people and society caused by counterfeit medicines,yet it is often slower to develop than counterfeit technology and is costly to detect.Therefore,inspection technology improvements are expected to be fast,non-destructive and low cost.Raman spectroscopy is based on the molecular level to meet the requirements of the pharmaceutical field for detection technology,becoming a popular analytical detection technology.In the last two years,attempts have been made to apply deep learning methods to Raman spectroscopy analysis techniques with good results.Raman analysis techniques and deep learning methods each have problems in their practical application.Such as:(1)Raman analysis requires spectral pre-processing,and the spectral data contains both linear and nonlinear factors,traditional pre-processing methods can usually only deal with one of the factors,making the spectra pre-processed results have a one-sided nature;(2)deep learning methods generally require a large number of data sets for training,and Raman spectra are usually not collected in large quantities,so deep learning applied to Raman analysis techniques face the problem of insufficient sample size of data sets;(3)existing analysis methods and models are for a single problem,few methods can simultaneously integrate pre-processing,classification and other related functions.To sum up,existing Raman analysis techniques and deep learning in Raman analysis are still shackled in their application.This study is based on the above problem to design the RAMAN-DCGAN model.(2)Deep learning methods generally require a large number of data sets for training,while Raman spectroscopy usually does not make use of the convolutional properties in large quantities to achieve the pre-processing of spectral data;secondly,the generation of adversarial networks to generate high-quality spectral data;finally,the convolutional layers in DCGAN are extracted to build new CNN models for spectral identification and classification.The study used Raman spectroscopic data from the China Institute of Food and Drug Validation to validate the model.The main results of this study are as follows.(1)Selection and improvement of pre-processing methods: Selection of convolution as a pre-processing method for spectral data,using the convolutional nature to blur unnecessary data in the spectrum,extract important data,achieve noise cancellation and baseline correction of the spectrum.(2)Data expansion method selection and improvement: DCGAN was selected as the data expansion method,and the convolutional layer in DCGAN was modified to apply to the Raman spectral data,so that DCGAN can generate higher quality Raman spectra to meet the needs of deep learning for training samples.(3)Model for migration learning: the convolutional layer in DCGAN is extracted and reconstructed into a convolutional neural network for the identification and classification of Raman spectra.The extracted convolutional layer is modified for Raman spectral properties and trained by DCGAN model,which greatly reduces the training time of the newly constructed CNN model.(4)Model validation: run using the NIFDC data entry model to verify the validity of the model.The raw data,DCGAN generation data and data generated by traditional data expansion methods are input into the traditional machine learning model and CNN model together,and the results are compared to verify the superiority of RAMAN-DCGAN model.