| Raman spectroscopy can detect the composition information of substances,and has many advantages such as fast analysis speed,low cost and simple operation.It has been widely used in agriculture,medicine,industry and many other fields,especially in the field of adulteration identification of traditional Chinese medicine.In this paper,adulteration identification of pearl powder of traditional Chinese medicine was studied based on Raman spectroscopy combined with deep learning algorithm.Firstly,the basic mechanism of Raman scattering and the basic theory of the currently commonly used chemometrics algorithm were introduced.Nacre powder and pearl powder were mixed in different proportions,and the simulated admixture of traditional Chinese medicine samples were prepared and the Raman spectra were collected.The spectra has been pretreated by methods including baseline correction and Savitzky-Golay smoothing filtering.The original Raman spectral data were enhanced by left-right translation and noise addition respectively,and the enhanced spectral data were normalized.Secondly,the qualitative identification of adulteration of pearl powder based on the deep learning network with different structures were studied.Four structures of cyclic neural network,convolutional neural network,cyclic network and convolutional network series mixed neural network and parallel mixed neural network were designed respectively.The network parameters were optimized to determine the optimal network structure,and compared with the traditional classification methods.The experimental results showed that convolution and cyclic parallel neural network had better anti-noise ability compared with other methods,which is suitable for the field rapid inspection of traditional Chinese medicine.Thirdly,this paper proposes a Raman spectral data enhancement algorithm based on deep convolution generative adversation network,and carries out spectral data enhancement on the real and false Raman spectra of pearl powder.The peak signal-to-noise ratio and structural similarity are used as evaluation indexes to compare with traditional data enhancement methods.Finally,the deep convolutional generative adversation network combined with convolutional neural network was used for qualitative and quantitative analysis of four kinds of low-purity adulterated pearl powder.Based on the qualitative classification model and quantitative model of the deep convolution generative antagonistic network enhanced spectrum,the results of adulteration identification are obviously better than the traditional methods.The experimental results prove that the enhancement of Raman spectral data based on deep convolution generative adconfrontation network can solve the problem of the lack of experimental Raman spectral data collected in many fields,and provide a new idea for the application of deep learning technology in the field of spectral analysis. |
