Research And Software Implementation Of Raman Spectral Component Identification Algorithm

Raman spectroscopy is a molecular vibrational spectroscopy,which is widely used in qualitative analysis in the fields of food safety,biomedicine,and medicinal chemistry,etc.due to its rapid and non-destructive characteristics.The primary goal of qualitative analysis is to correctly identify components.However,due to environmental factors,Raman spectroscopy is prone to some nonlinear changes,such as spectral peak drift,broadening or distortion,making component identification difficult.In addition,over time,due to instrument aging and other reasons,the spectrum acquired in tihe past may no longer meet the current statistical distribution,resulting in the original recognition algorithm failure.In order to solve these problems,this paper explores the relevant methods of spectral component identification,and writes the supporting online identification software.The main work of this paper is as follows:A single spectral denoising method based on convolutional denoising autoencoder is proposed.The method firstly injects noise into the spectrum,obtains training data with high and low SNR matching,and then obtains a convolutional denoising autoencoder for the spectrum with greedy layer-wise unsupervised pretraining.Finally,we used it for prediction to reach the goal of denoising.This method can denoising automatically,and the performance is better than the traditional algorithm.A multi-spectral denoising method based on singular value decomposition and median absolute deviation is proposed.The method firstly performs singular value decomposition on multiple spectra,and then uses the median absolute deviation method to screen out the top k singular values with outlier characteristics.Finally,we use these singular values to re-solve the denoised spectrum.This method is suitable for the processing of hyperspectral data and has a great improvement in image quality.An online Raman spectral component identification method based on multitask convolutional base and non-shared convolutional neural network is proposed.The method firstly obtains the multitask convolutional base through the uniform data set,then freezes and transfers it to the non-shared convolutional neural network for identifying the components of interest to obtain the initial model.Finally,the model is continuously tuned in line with the way of online learning,and the identification performance of the model is improved.This method has a good generalization ability on small batch samples,and with the help of online learning,it also overcomes the timeliness of spectral data.The Raman spectral component identification software based on front-end separation technology was written.The software is written in TypeScript and Python and implements all of the algorithms mentioned in this article and some common spectral preprocessing algorithms.