An Artificial Neural Network to Eliminate the Detrimental Spectral Shift on Mid-Infrared Gas Spectroscopy

Abstract

We demonstrate the successful implementation of an artificial neural network (ANN) to eliminate detrimental spectral shifts imposed in the measurement of laser absorption spectrometers (LASs). Since LASs rely on the analysis of the spectral characteristics of biological and chemical molecules, their accuracy and precision is especially prone to the presence of unwanted spectral shift in the measured molecular absorption spectrum over the reference spectrum. In this paper, an ANN was applied to a scanning grating-based mid-infrared trace gas sensing system, which suffers from temperature-induced spectral shifts. Using the HITRAN database, we generated synthetic gas absorbance spectra with random spectral shifts for training and validation. The ANN was trained with these synthetic spectra to identify the occurrence of spectral shifts. Our experimental verification unambiguously proves that such an ANN can be an excellent tool to accurately retrieve the gas concentration from imprecise or distorted spectra of gas absorption. Due to the global shift of the measured gas absorption spectrum, the accuracy of the retrieved gas concentration using a typical least-mean-squares fitting algorithm was considerably degraded by 40.3%. However, when the gas concentration of the same measurement dataset was predicted by the proposed multilayer perceptron network, the sensing accuracy significantly improved by reducing the error to less than ±1% while preserving the sensing sensitivity.