Methane detection to 1 ppm using machine learning analysis of atmospheric pressure plasma optical emission spectra

Abstract

Optical emission spectroscopy from a small-volume, 5 μl, atmospheric pressure RF-driven helium plasma was used in conjunction with partial least squares-discriminant analysis for the detection of trace concentrations of methane gas. A limit of detection of 1 ppm was obtained and sample concentrations up to 100 ppm CH4 were classified using a nine-category model. A range of algorithm enhancements were investigated including regularization, simple data segmentation and subset selection, feature selection via Variable Importance in Projection and wavelength variable compression in order to address the high dimensionality and collinearity of spectral emission data. These approaches showed the potential for significant reduction in the number of wavelength variables and the spectral resolution/bandwidth. Wavelength variable compression exhibited reliable predictive performance, with accuracy values >97%, under more challenging multi-session train—test scenarios. Simple modelling of plasma electron energy distribution functions highlights the complex cross-sensitivities between the target methane, its dissociation products and atmospheric impurities and their impact on excitation and emission.