摘要
电力智能巡检与边缘计算相结合,对电力物联网和透明电网的建设具有重要作用。然而,边缘设备的低算力导致检测算法在边缘端推理速度慢,边缘设备的低内存也限制了目标检测模型的空间占用。针对以上问题,提出了一种基于重参数化YOLOv5的输电线路缺陷边缘智能检测方法。首先,利用R-D模块和重参数化空间金字塔池化(spatial pyramid pooling,SPP)对YOLOv5网络进行改进,通过重参数化加快模型的推理速度,得到重参数化YOLOv5模型。其次,利用ResRep剪枝方法对模型进行通道剪枝,减小模型空间占用。最后,将模型部署至NVIDIA Jetson Xavier NX边缘平台,并利用C++语言结合TensorRT图优化对模型进行优化加速,进一步减小推理时延,减少模型占用空间。实验结果表明:相较于原版YOLOv5,该文提出的方法推理速度提升至5倍,空间占用减小至41%,同时精度提升了2.4%,显著提高了输电线路边缘智能巡检的效率。
The combination of power intelligent inspection and edge computing plays a crucial role in the construction of power Internet of Things and transparent power grid.However,the low computing power of edge devices leads to the low speed of models running on edge devices,and the low memory of edge devices also limits the memory usage of object detection models.Aiming at the above problems,we put forward a method for transmission line defect edge intelligent inspection based on re-parameterized YOLOv5.First,R-D block and re-parameterized spatial pyramid pooling(SPP)are utilized to improve YOLOv5,which adopt the re-parameterization technology to enhance the inference speed,and re-parameterized YOLOv5 is developed hereby.In addition,ResRep is used to perform channel pruning on the model,decreasing the memory usage of the model.Finally,the model is deployed to NVIDIA Jetson Xavier NX embedded platform,and C++language combined with TensorRT graph optimization is employed to optimize and accelerate the model,further reducing the inference latency and memory usage.The experimental results show that the proposed method is five times the inference speed and forty-one percent the memory cost of YOLOv5,with accuracy being improved by 2.4%,therefore,the efficiency of transmission line edge intelligent inspection can be significantly improved.
作者
刘闽
李喆
李曜丞
刘亚东
江秀臣
LIU Min;LI Zhe;LI Yaocheng;LIU Yadong;JIANG Xiuchen(Department of Electrical Engineering,Shanghai Jiao Tong University,Shanghai 200240,China)
出处
《高电压技术》
EI
CAS
CSCD
北大核心
2024年第5期1954-1966,共13页
High Voltage Engineering
基金
南方电网公司科技项目(GDKJXM20220236)。
