Machine Learning Algorithm Improved Qualitative And Quantitative Ability Of Raman Spectroscopy

In recent years,the continuous upgrading of analytical instruments and related technologies has put forward higher demands for data analysis methods.In this paper,how to achieve fast and convenient Raman spectroscopy detection as the research goal,we attempted to improve the qualitative and quantitative analysis ability of Raman spectroscopy technology by machine learning.The main work is divided into three parts:(1)Development of denoising algorithms to improve temporal resolution.High temporal resolution is of great significance to the research of the dynamic process of interfacial catalysis and biological reaction,and the key to improve the temporal resolution lies in the extraction of weak signals.Therefore,we develop the peak extraction and retention denoising algorithm(Peak Extraction and Retention,PEER).Compared with the Savitzky-Golay and wavelet denoising algorithms,the PEER algorithm can retain the signal with high SNR,and realize the effective extraction of the weak signal without anti-background interference.Further,compared with the fast Fourier transform algorithm,the PEER algorithm improves clarity and temporal resolution of Raman imaging.For example,the temporal resolution of Raman imaging in living cells can be increased by an order of magnitude.(2)Development of qualitative analysis methods for trace mixtures.The SERS intensity changes caused by the partial overlap of peaks of each target and the competitive adsorption between the targets will significantly increase the difficulty of qualitative analysis of the target.In this chapter,we try to improve the accuracy and repeatability of qualitative analysis by combining data augmentation and random forest.The data augmentation provides sufficient data for training model,and the then random forest can realize fast and high accurate qualitative analysis from the perspective of data analysis.In addition,we also visualize the random forest to enhance interpretation of the model.Futher,the model is optimized through two aspects:accuracy and visualization.By using the qualitative analysis of three trace PAHs mixtures as the model system,the reliable qualitative analysis of these three trace PAHs can be realized by means of the SERS spectrum of the target and the visual analysis of the random forest(3)Development of quantitative analysis methods for salinity oriented to needs.The non-contact and information-rich Raman spectroscopy has the potential to solve the problems of susceptibility to seawater corrosion and large error on the salinity meter.However,there are still large errors in traditional analysis methods(such as using different peak strength to realize quantitative analysis).Therefore,by introducing principal component analysis,the effective information is extracted more completely.The quantitative analysis of salinity is realized by combining polynomial regression algorithm.Compared with the traditional,the method has significant advantages in error,repeatability and fitting degree of the model.Compared with commercial salinity meter,the method can be used in seawater with different components and has advantages in general use.