Multi-Gas Regression from High-Speed Image Sequences Using 3D CNN and 3DResNet Architectures in Biomass Co-Combustion: A Feasibility Case Study

This study explored a spatio-temporal deep learning approach for optical soft sensing of combustion emissions in a coal–biomass co-firing scenario. High-speed RGB flame sequences from a 0.5 MW test rig co-firing hard coal with 10% straw were synchronized with extractive measurements of O2, CO2, and NO. These sequences were used to train three shallow 3D CNNs and three 3D ResNet-50 architectures with squeeze-and-excitation attention. The proposed 3D CNN/ResNet models performed simultaneous regression of all three gas species from flame image volumes. The best configuration achieves R2 values of 0.975, 0.987, and 0.980, accompanied by mean absolute errors of 0.23% by volume, 13.15 mg/m3, and 0.19% by volume for O2, NO, and CO2, respectively, at a resolution of 128 × 96 × 96 pixels. Within the scope of the available dataset, comprising a single measurement run and a single fuel mixture, the results indicate that a comprehensive spatio-temporal analysis of flame images can yield accurate estimates of multiple gas concentrations, thereby providing a promising foundation for the future development of soft optical sensors. At the same time, the study is limited to a single combustion experiment, a single biomass fraction, and a single borescope orientation, and the inference delay and hardware requirements were not quantified; therefore, issues regarding the generalizability of the proposed approach to different conditions and its implementation remain open for further work.