ConInceDeep: A novel deep learning method for component identification of mixture based on Raman spectroscopy

For mixture component identification, the methods based on deep learning are becoming prevalent due to their end-to-end characteristic, being completely data-driven and reducing the demand for prior knowledge. The most used CNN model currently, however, relies only on the original one-dimensional (1D) spectral data. When faced with complex mixtures, the features of overlapping peaks and weak peaks are probably not sufficient mined by 1D CNN model, which would reduce the identification accuracy. Thus, a novel deep learning method entitled ConInceDeep for component identification of mixture by Raman spectroscopy was proposed in this study. The ConInceDeep firstly performed continuous wavelet transform on 1D spectral data using the Lorentz4 wavelet to fully reveal the detailed characteristics of weak and overlapping spectral peaks in the mixture and then employed Inception modules consisting of multiple-size kernels to construct two-dimensional CNN model to improve the model's adaptability to different Raman peaks. The proposed ConInceDeep model was trained and validated entirely by the spectra of virtual mixtures which were generated by simple manipulation of substances' spectra in the database. Three mixture datasets were used for verifying the identification performance of the ConInceDeep, including 191 liquid mixtures, 33 powder mixtures and two kinds of real samples. In liquid mixture set, ConInceDeep was compared with other three experimental methods, and achieved an increase of Acc from 90.32% to 96.60% compared with the traditional 1D-CNN. In powder mixtures set, ConInceDeep reached 99.66% Acc and 0.35% FPR by 54 models. Additionally, it was also applied to real samples to demonstrate its value. Summarily, the ConInceDeep provides a new reference for mixture identification based on Raman spectroscopy.