基于注意力深度神经网络的电力场景火焰烟雾视频监测  

Video Monitoring of Flame Smoke in Electric Power Scenes Based on Attention Deep Neural Network

作　　者：马洲俊 陈锦铭 赵晨萌 张卡[3] Ma Zhoujun;Chen Jinming;Zhao Chenmeng;Zhang Ka(State Grid Jiangsu Electric Power Co.,Ltd.,Nanjing 210019,China;Electric Power Research Institute,State Grid Jiangsu Electric Power Co.,Ltd.,Nanjing 211103,China;Nanjing Normal University,Nanjing 210023,China)

机构地区：[1]国网江苏省电力有限公司,南京210019 [2]国网江苏省电力有限公司电力科学研究院,南京211103 [3]南京师范大学,南京210023

出　　处：《科技通报》2025年第2期8-14,38,共8页Bulletin of Science and Technology

基　　金：国网江苏省电力有限公司科技项目(J2023121)。

摘　　要：针对电力场景中火焰烟雾等非刚体目标的预警响应需求,本文提出了融合注意力机制和深度学习的视频火焰烟雾自动监测算法。首先,以YOLOv8为主干网络,引入三分支结构的无参注意力机制和内嵌注意力机制的动态目标检测头部框架,以改善由非刚体目标动态变化大且电力场景图像背景复杂等因素造成的目标遮挡和特征不明显等问题;然后,利用基于辅助边框的损失函数解决YOLOv8模型中信息流失和不充分整合的问题;最后,使用自制数据集fireANDsmoke对本文算法进行实验验证。实验结果表明:对于电力场景的监控视频数据,与原始YOLOv8相比,本文提出的火焰烟雾检测算法具有更高的检测准确性和鲁棒性,在mAP50与mAP50-95这2个指标上都提升了2.9%。本文的研究为电力场景火焰烟雾的自动监测提供了一种可行的方法,对于提高电网安全监控水平和应急响应能力等具有重要的应用价值。To address the need for early warning and response to non-rigid body targets such as smoke and flame in electric power scenarios,this paper proposes an automatic video monitoring algorithm for flame and smoke detection,integrating attention mechanisms with deep learning.Initially,YOLOv8 is employed as the backbone network,enhancements include the introduction of a nonparametric attention mechanism with a three-branch structure and a dynamic target detection head embedded with attention mechanisms,these improvements mitigate issues of target occlusion and feature obscuration caused by the dynamic nature of non-rigid objects and the complex backgrounds in power scene images;Additionally,an auxiliary edge-based loss function is utilized to address the information loss and integration deficiencies in the original YOLOv8 model.The proposed algorithm is validated experimentally using a custom dataset named fireANDsmoke.The experimental outcomes confirm that the introduced algorithm secures an enhancement in detection precision and robustness compared to the original YOLOv8 when applied to surveillance video data in power scenarios,with an improvement of 2.9%in mAP50 and mAP50-95 metrics.The paper’s research provides a feasible method for automatic monitoring of smoke and flame in electric power scenes,which has important application value in improving security monitoring level and emergency response ability for the power grid.

关 键 词：电力场景 火焰烟雾 非刚体目标检测 注意力机制 深度学习 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]