Single convolutional neural network model for multiple preprocessing of Raman spectra

Abstract

Raman spectra suffer from the interferences of noises, baseline drifts, and cosmic rays, which leads to errors in the subsequent analysis of the spectra. The commonly applied spectral preprocessing methods, such as wavelet transform (WT), Savitzky-Golay smooth (SG smooth), asymmetric least squares (AsLS) can only correspondingly reduce a single interference term, and the completion of preprocessing based on these traditional methods requires a series of tedious trials. Especially, each scheme can only be used for a specific data set. Convolutional neural network (CNN) is commonly used for object recognition, image super-resolution, and natural language processing. CNN has the potential to be applied to the preprocessing of chemical signals to solve problems intelligently. In this paper, we developed a method to remove all the interferences from multiple Raman spectra data sets collected from food and wastewater samples, using only a single CNN model. By optimization of hyperparameters, activity functions, and loss functions in CNN, the CNN model completely removed noises, spikes, baselines, and cosmic rays from Raman spectra and simplified the preprocessing of Raman spectra.