Region-aware multimodal deep feature fusion of magnetically confined premixed ethylene micro-jet flames

Reliable diagnosis of subtle magnetic-field-induced changes in fuel-rich micro-jet flames remained difficult when flame morphology, radical emission, and spectral response were interpreted separately. This study investigated premixed ethylene (C 2 H 4 ) micro-jet flames under magnetic confinement and developed a region-aware multimodal information-fusion approach for flame-state diagnosis. Flame photography, OH*/CH* chemiluminescence imaging, emission spectroscopy, and temperature measurements characterized the magnetic flame response. The magnetic confinement shortened the flame, increased its width, and produced a denser luminous structure. The flame temperature increased at all measured heights, while the optical response showed slightly increased OH* emission and slightly decreased CH* and C 2 * related spectral features, indicating a more compact and effective oxidation process. Because spectral differences between magnetic states were small relative to flame-to-flame variability, spectroscopy alone provided limited classification ability. At φ = 2.00, the best individual modality, CH* imaging, achieved a mean test accuracy of 78.67 %. Region-aware CNN-based Xception and MLP deep feature extraction for Color, CH*, OH*, and spectral information were fused and classified by Ensemble Tree; the mean test accuracy increased to 90.33 % (+11.66 percentage points) with balanced precision, sensitivity, specificity, and F1 score. These results demonstrate that magnetic flame states are best represented through complementary spatial and spectral observables.