A TFA-CNN method for quantitative analysis in infrared spectroscopy

Abstract

Infrared spectroscopy has been widely used in quantitative analysis due to its advantages of rapidity, high efficiency and no pollution. With the development of deep learning, Convolutional Neural Network (CNN) promotes the further development of quantitative analysis methods. Spectral pretreatment is a key part to remove noise before the network training, which is able to improve the accuracy of the CNN model. However, the parameters adjustment and the signal information loss always exist in the preprocessing, which limits the application scope of CNN model. In this paper, a combined Time-FrequencyAnalysis and CNN (TFA-CNN) method for quantitative analysis in infrared spectroscopy is proposed. The pretreatment process of infrared spectra is integrated with CNN to avoid the loss of information, guaranteeing a high accuracy quantitative analysis without parameter adjustment. The experiment results show that the proposed TFA-CNN method has better accuracy in quantitative analysis of infrared spectroscopy, compared with traditional quantitative analysis methods. Furthermore, the TFA-CNN method has proved to be effective in a relatively small dataset. The proposed method is simple and universal, which has a broader application prospect.