Bio-FlameNet: Flame Detection for Biogas Digesters

Biogas facilities are crucial for converting organic waste into renewable energy, yet their inherent methane-related fire and explosion risks pose significant operational safety threats. Deploying automated early-flame detection systems in this high-risk industrial setting faces two core challenges. First, there is extreme scarcity of domain-specific real-world fire data due to safety constraints. Second, the detector must alarm at the incipient stage of a fire, where the target is extremely small. This paper proposes Bio-FlameNet, a data-centric framework for building an efficient and reliable flame detector in the absence of real-world fire data. We generate a high-fidelity synthetic dataset by fusing flame assets onto real biogas-facility backgrounds using a copy–paste pipeline, explicitly controlling the small:medium:large flame ratio to 7:2:1. We further adopt a two-stage transfer-learning strategy by pre-training on D-Fire and fine-tuning on domain-specific synthetic data. Experiments on the YOLOv8 family show that the proposed approach improves lightweight-model performance while substantially reducing training time across all model scales.