The paper explores the potential use of neural networks to predict the chemical composition as well as the total and partial pressure of individual gases in the mixture based on their emission spectra. Excitation/ionization of gas particles is ensured by a miniature MEMS (micro-electro-mechanical) sensor. The impact of measurement conditions on the generated spectra and possible challenges in analysis are addressed. The trained neural networks achieved a mean absolute error of approximately 1.4% in determining the chemical composition of a mixture of methane and hydrogen and 0.8% for a mixture of nitrogen and oxygen. For pressure prediction, spanning several orders of magnitude, a general model predicting pressure for ten different gases achieved a relative mean absolute error of 4.5%. Additionally, a model for gas classification achieved an accuracy of over 99%.
Read full abstract