基于边缘智能的煤矿外因火灾感知方法  

Coal mine external fire detection method based on edge intelligence

作　　者：赵端[1,2] 李涛 董彦强 王志强 刘春 ZHAO Duan;LI Tao;DONG Yanqiang;WANG Zhiqiang;LIU Chun(National and Local Joint Engineering Laboratory of Mining Internet Application Technology,China University of Mining and Technology,Xuzhou 221116,China;School of Information and Control Engineering,China University of Mining and Technology,Xuzhou 221116,China;Inner Mongolia Baiyinhua Coal and Electricity Co.,Ltd.,State Power Investment Group,Xilin Gol 026200,China)

机构地区：[1]中国矿业大学矿山互联网应用技术国家地方联合工程实验室,江苏徐州221116 [2]中国矿业大学信息与控制工程学院,江苏徐州221116 [3]国家电投集团内蒙古白音华煤电有限公司,内蒙古锡林郭勒026200

出　　处：《工矿自动化》2022年第12期108-115,共8页Journal Of Mine Automation

基　　金：中央高校基本科研业务费专项资金项目(XJ2021005101)。

摘　　要：对煤矿外因火灾隐患进行检测,实现对初期火灾的可靠判识,对于提升煤矿火灾检测水平有重要意义,也是未来智能矿山建设的重要方向。为了提高煤矿外因火灾检测速度、精度和实时性,提出一种基于边缘智能的煤矿外因火灾感知方法。对YOLOv5s模型主干网络特征尺度进行改进,使模型能够充分学习浅层特征,改善小目标检测性能,同时在原有的特征金字塔网络(FPN)基础上添加自适应注意模块,提高模型检测精度。为解决井下光线条件差、粉尘多及摄像机拍摄角度引起的图像检测误差和漏检问题,采用多传感器辅助检测,通过动态加权算法对视频检测信息和多传感器检测信息进行加权融合判识,构建了YOLOv5s−as模型。将YOLOv5s−as模型移植到智能边缘处理器上,并进行轻量化处理,实现边缘智能设备部署。实验结果表明:与未加入传感器信息融合推理的YOLOv5s−a模型相比,YOLOv5s−as模型推理时间略长,但交并比为0.5时的平均精度均值(mAP@0.5)提高了7.24%;与移植前的YOLOv5s模型相比,移植到智能边缘处理器上并进行轻量化处理的YOLOv5s−as模型mAP@0.5提高15.04%;SSD 300,SSD 512及YOLOv5s模型无法识别小目标火源,YOLOv5s−a,YOLOv5s−as模型能够检测出小目标火源,适应性较好;使用边缘处理方式时,YOLOv5s−as模型的响应周期为238 ms,比集中式处理方法缩短了38.66%。The detection of external fire in coal mines and the reliable identification of initial fire are of great significance for improving the level of coal mine fire detection.It is also an important direction of intelligent mine construction in the future.In order to improve the speed,precision and real-time of coal mine external fire detection,a coal mine external fire detection method based on edge intelligence is proposed.The feature scale of the backbone network of the YOLOv5s model is improved.The model can fully learn the shallow features and improve the small target detection performance.At the same time,an adaptive attention module is added on the basis of the original feature pyramid network(FPN)to improve the detection precision of the model.There are problems of image detection error and missed detection caused by poor light conditions,more dust and camera shooting angle in the underground mine.In order to solve the above problems,the YOLOv5s-as model is constructed by using multi-sensor auxiliary detection and weighting fusion identification of video detection information and multi-sensor detection information through dynamic weighting algorithm.The YOLOv5s-as model is transplanted to the intelligent edge processor,and lightweight processing is carried out to realize the deployment of edge intelligent devices.The experimental results show that the reasoning time of the YOLOv5s-as model is slightly longer than that of the YOLOv5s-a model without sensor information fusion reasoning,but mean value of average precision when the intersection over union is 0.5(mAP@0.5)is increased by 7.24%.Compared with the YOLOv5s model before transplantation,the mAP@0.5 of the YOLOv5s-as model transplanted to the intelligent edge processor and subjected to lightweight processing increased by 15.04%.For small target fire sources,SSD 300,SSD 512 and YOLOv5s models cannot identify them.The YOLOv5s-a and YOLOv5s-as models can detect small target fire sources with good adaptability.When using the edge processing method,the respons

关 键 词：外因火灾检测 YOLOv5s 深度学习 自适应注意机制 多源信息融合 边缘计算 

分 类 号：TD752.3[矿业工程—矿井通风与安全]