Few Shot Learning for Flame State Monitoring From Limited Visible and Infrared Images

ABSTRACT The success of machine learning in image‐based combustion monitoring requires big data, which is costly, and even impossible in industrial applications. To address this, we introduce metric‐based few shot learning for combustion monitoring and classification, to the best of our knowledge, for the first time. Two algorithms, that is, Siamese Network coupled with k Nearest Neighbors (SN‐kNN) and Prototypical Network (PN), are proposed. We explore the use of both visible and infrared (IR) images. We analyse the training process, test performance and inference speed of two algorithms using these image modalities, while t‐SNE is used to visualize the learned features. The results demonstrate that both SN‐kNN and PN are capable of distinguishing flame states, by just using 20 images per state. It is found that using visible images offers the ability to distinguish more substantial differences between classes, and identify more consistent patterns within each class, leading to higher training speed and model performance. In contrast, the relatively low quality of IR images complicates the extraction of distinguishable prototypes, as is in the case of PN, however, nevertheless offering high classification rates (F1‐score 0.99 averaged over five random seeds). When using the entire training set, SN‐kNN performed well with IR images. On the other hand, due to its architectural design, PN offers faster training speed and inference. To the best of our knowledge, this is the first application of few shot learning methods on both image modalities for combustion state monitoring. We further position few shot learning among alternative data‐efficient learning paradigms, analyse the practical deployment trade‐offs of the two models, provide a controlled ablation of the principal design choices, and quantify temporal generalization, showing that a temporally representative 20‐image calibration set restores near‐ceiling accuracy under a strict temporal split, thus offering guidance for their future utilization in combustion monitoring tasks.