One Dimensional Convolutional Neural Networks for Spectral Analysis

Abstract

A platform for building sensor specific machine learning detection algorithms has been developed to classify spectroscopic data. The algorithms are focused on long wave infrared reflectance (LWIR) and Raman spectroscopies. The classification algorithm is based on a one dimensional (1D) convolutional neural network (CNN) architecture. Training data is generated using an appropriate signal model that is combined with sensor specific characteristics such as spectral range, spectral resolution, and noise. Within this paper, the performance of trained CNNs for both LWIR and Raman sensor systems has been evaluated. The evaluation uses both real and synthetic data to benchmark the performance in terms of the discriminant signal. The evaluation data consists of various chemical representations and varied noise levels. The performance of the 1D CNN approach has demonstrated high classification accuracies on data with low discriminant signals. Specifically, the CNNs have demonstrated a classification accuracy >90% for infrared reflectance data down to a wavelength averaged discriminant SNR>1. For Raman systems, we have demonstrated classification accuracies >90% for data with a peak discriminant SNR of approximately 6.