Power Material Recognition Based On Infrared Spectroscopy And Machine Learning

Infrared spectroscopy can effectively carry information about the structure of compounds and their constituent components.Its applications in chemical research,purity testing and drug identification have been widely achieved.However,in practical application scenarios,this technique is also greatly limited by the lack of standard sample preparation conditions,the low accuracy of infrared spectroscopy identification and the poor identification efficiency.In this paper,a tunable infrared quantum cascade laser is used as the laser source to establish an experimental system for recording the reflected and diffuse reflectance spectra of powder samples,combined with a neural network machine learning algorithm for the non-destructive and rapid detection and analysis of powder samples without preparation.The following work in progress is presented in this paper:1.Infrared reflectance spectra of glucose samples were obtained in the reflectance optical path to calculate the transmittance of powder samples at different concentrations.It was found that the samples showed absorption peaks at the same position at different concentrations,which can prove that the infrared spectra carry information about the structure of the samples and that the intensity of the transmittance decreases with increasing concentration,which can be used for concentration specific analysis in the next work.2.The IR diffuse reflectance spectra of the glucose samples were obtained in the diffuse reflectance optical path,and the Kubelka-Munk equation and the Kramer-Kronig relationship were applied to the experimental results to synthesise the transmission spectra of the samples,and the possibility of the diffuse reflectance spectra reducing the transmission spectra was verified by comparison with the IR spectra in the database.3.The diffuse reflectance spectral data were applied to two neural network models for the prediction of mixed powder mass fractions.Under the Kramers-Kronig relational transformation,the LSTM network model predicted the best results,significantly better than the BP neural network model.Under the Kubelka-Munk equation transformation,both neural networks were more accurate in predicting low mass fraction polyethylene samples and poorer in predicting high mass fractions.Both diffuse reflectance spectral correction methods improved the accuracy of the training results to varying degrees,with the LSTM model outperforming the BP neural network model overall.These findings contribute to the development of techniques for the identification of unknown mixed powder samples based on frequency-tunable or broad-spectrum infrared lasers.