Real-time analysis of flame chemiluminescence spectra for equivalence ratio and gas composition using neural network approaches

In response to the expected challenges in gas network operations, where variable gas compositions require robust combustion systems with automatic adaptation, this study explores the application of an artificial neural network (ANN) for real-time recognition of flame feed gas compositions based on spatially integrated flame chemiluminescence in the 285–550 nm spectral range. Operating conditions of a premixed laminar jet flame were dynamically varied in respect of equivalence ratio (ϕ), fuel composition (methane/hydrogen blend), and inert gas composition (carbon dioxide/nitrogen blend). Out of the >300 conditions investigated, 80 % served as training data and the remaining 20 % were analyzed to demonstrate the predictive accuracy. The overall mean absolute error ($MAE$) relative to the ground truth data for ϕ was 0.016, for the methane/hydrogen blend 1.62 %, and for the inert gas variation 0.3 %. To further test the networks performance, the network was also applied to randomly selected spectra from a second dataset acquired under slightly different conditions in respect of gas exit velocity and background illumination. This analysis also achieved low $MAEs$ of 0.04 for ϕ, 4.6 % for the methane/hydrogen, and 0.49 % for inert gas variation. The standardized absolute deviation ($SAD$), showed that there is no correlation between individual operating conditions and the related errors in the data analysis. An analysis time of 10 ms allows the network to be used for real-time application.