Flame Detection in Converter Stations Using Enhanced U-Net and HyCTAS with Binocular Vision

Catastrophic substation accidents are usually accompanied by fires and explosions. Fire events and its position detection around power transformers arouse increasing attention, especially in converter stations equipped with soundproof covers. This paper introduces a flame detection framework based on binocular cameras. The proposed framework incorporate an enhanced U-Net model with ResNet50, HyCTAS self-attention mechanisms, and binocular vision-based 3D localization. By improving attention to complex flame patterns with HyCTAS, this model significantly outperforms baseline methods by 15% increase in detection precision and 10% increase in recall value. The onsite deployed binocular cameras are mainly used for flame locating, which benefits accurate rescue from outside the soundproof covers. This work suggests significant potentials of AI-based models in the field of fire monitoring and mitigation in converter stations. Future research will focus on computational costs reduction, model optimization and onsite applications.