Abstract
Reliable fire detection is essential for ensuring the safe operation of power distribution substations. This study proposes a fire recognition framework for substation surveillance environments. First, an illumination adaptation module is introduced to enhance feature representation under varying illumination conditions. Second, an adaptive spatial feature extraction framework is developed to jointly capture local smoke patterns and large-scale diffusion structures. Third, an illumination-aware classification loss is proposed to explicitly incorporate illumination information into the optimization process, thereby improving recognition performance under challenging illumination conditions. Experimental results demonstrate that the proposed framework achieves superior recognition performance compared with representative vision models while maintaining competitive inference efficiency, highlighting its practical applicability to substation surveillance.