【作者】 王卓；

【导师】 康守强；

【作者基本信息】 哈尔滨理工大学 ， 电子与通信工程（专业学位）， 2021， 硕士

【摘要】 绝缘子作为电力系统中重要的绝缘装置,在架空输电线路中起着机械支撑和防止电流回地的重要作用。受灾害、温度、潮湿等自然因素的影响,绝缘子很容易出现爆缸、金属护具脱落等物理故障,一旦绝缘子发生故障,直接威胁着整个电力系统的输电稳定性,产生严重的安全隐患和经济损失,因此,电力系统中绝缘子的状态监测一直备受关注。随着无人机设备的进步,基于航拍图像的输电线路中关键电气设备的故障巡检方式,已经成为线路巡检的研究热点。本文围绕深度学习算法,对航拍图像的输电线路中绝缘子定位和故障检测进行研究,针对目标检测算法易受复杂背景干扰,导致准确率低的问题,提出一种基于协同深度学习的二阶段绝缘子故障检测方法,具体流程如下:(1)第一阶段,利用FCN算法对航拍图像预处理,设计跳跃结构融合浅层图像特征与深层语义特征,构建8倍上采样的绝缘子分割模型,初步分割出绝缘子区域图像,然后将分割后的绝缘子区域图像与原图像进行逻辑运算处理,得到去除背景区域的绝缘子图像;(2)第二阶段,将去除背景区域的绝缘子图像作为训练集数据,构建针对绝缘子故障的YOLOv3目标检测模型,以基于Darknet-19网络改进的Darknet-53作为特征提取网络,并结合特征金字塔思想,在输出张量的3个尺度上对绝缘子故障区域进行标记和类别预测。针对YOLOv3网络的默认锚点框参数对绝缘子数据集的不匹配问题,利用K-means++聚类算法优化YOLOv3的锚点框参数,进一步提升检测精度。实验结果表明,基于协同深度学习的二阶段方法能够有效克服复杂背景的干扰,在绝缘子故障检测中平均准确率(Mean Average Precision,MAP)高达96.88%,较原始YOLOv3算法MAP值提升了4.65%。更多还原

【Abstract】 As an important insulation device in power system,insulators play an important role in mechanical support and preventing current from returning to ground in overhead transmission line,Affected by natural factors such as disaster,temperature and humidity,the insulator is prone to detonation cylinder,metal gear fall off and other physical faults,once the insulator fails,it will directly threaten the transmission stability of the whole power system,and cause serious security hidden danger and economic losses,therefore,in the power system of insulator state monitoring has been closely concerned.With the rapid development of UAV technology,the fault inspection of key electrical equipment in transmission lines based on aerial image has become a research hotspot.Focusing on the deep learning algorithm,this paper studies insulator location and fault detection in transmission lines of aerial images.Aiming at the problem that target detection algorithm is vulnerable to complex background interference,leading to low accuracy,a two-stage insulator fault detection method based on collaborative deep learning is proposed,the specific process is as follows:(1)The first stage,the FCN algorithm is used to preprocess the aerial image.The jump structure is designed to fuse the shallow image features and deep semantic features,and an 8-fold up-sampling insulator segmentation model is constructed,the insulator region image is preliminarily segmented,then the segmented insulator region image and the original image are logically processed to get the insulator image with the background region removed;(2)The second stage,the insulator image with the background region removed is used as the training set data,construct the insulator fault YOLOv3 object detection model,the improved Darknet-53 based on Darknet-19 network is used as the feature extraction network,and combining with the idea of feature pyramid,the insulator fault regions can be marked and classified on three scales of the output tensors.For the mismatch problem of default anchor frame parameters of YOLOv3 network to insulator data set,K-means++ clustering algorithm is used to optimize the anchor boxes parameters of YOLOv3 to further improve the detection accuracy.The experimental results show that the two-stage method based on collaborative deep learning can effectively overcome the interference of complex background,the mean average precision of insulator fault detection is as high as 96.88%,which is 4.65%higher than the original YOLOv3 algorithm.更多还原