Spatiotemporal Prior-Enhanced Fire Perception Network for Edge Alarms in Industrial LANs

Abstract Multi-camera fire monitoring in industrial local area networks is a challenging research problem requiring low-latency alarms under stringent uplink bandwidth constraints. Due to intense lighting and specular reflections, conventional edge vision models struggle with false positives, triggering unnecessary uploads of invalid region-of-interest ROI evidence. To address this, this empirical study improves the YOLO baseline by introducing a Prior-Aware Gating Block, PAGB, into the Backbone, which suppresses background noise and generates minimal semantic ROI evidence. Additionally, to mitigate strong-light false alarms without misfiltering steady-state flames, a Prior-guided Kinematic Attention Block, PKAB, is embedded into the Neck to capture short-term flicker and non-rigid motion cues via a negative-bias dual-branch fusion mechanism. Compared with the conventional YOLO baseline, the proposed PGENet achieves a Fire single-class \((\mathrm{mAP}@0.5)\) of 0.963 and, on the dedicated high-exposure interference subset, reduces the false-positive event rate from 36.2% to 2.1% at 32 FPS on the Huawei Atlas 200 DK. Under a detection-triggered edge architecture, PGENet reduces the measured single-alarm payload from 152 KB to 113 KB. Edge-device measurements show improvements in both detection accuracy and communication payload.