【作者】 徐杰；

【导师】 高丙团； 仲林林； 陈昊；

【作者基本信息】 东南大学 ， 电气工程（专业学位）， 2021， 硕士

【摘要】 我国南方沿海地区台风多发,破坏力强,极大地影响了电网的安全稳定。当电网遭遇台风、泥石流等灾害时,输配电线路的杆塔容易出现倾斜、倒塌等情况,极易导致线路出现故障,进而引发停电事故。为了快速、准确定位异常杆塔,及时恢复电力供应,电力运维部门通常使用无人机进行灾后电力杆塔巡检。然而,当前无人机巡检产生的海量巡检图像仍主要依靠人工处理,不仅费时费力,而且容易出现漏判和错判的情况。因此,本文采用基于深度学习的图像识别技术对线路杆塔进行智能化检测,并利用无人机巡检图像的元数据,通过坐标变换的方式计算获得线路杆塔的GPS信息,从而为灾后线路杆塔的快速抢修提供辅助决策信息,提高线路抢修能力。本文的主要研究内容如下:首先,研究了适用于无人机线路杆塔巡检图像的图像增强技术,构建了类别均匀、数量适中的无人机线路杆塔巡检图像数据库。针对无人机巡检拍摄时因曝光不足导致图像较暗的问题,使用灰度均衡方法,提升了图像质量;为了提高模型在各种环境下的图像识别鲁棒性,依据邻域风险最小化原则(Vicinal Risk Minimization,VRM),使用扭曲、滤波等图像增强方法,扩展了图像样本;为解决线路杆塔在整张巡检图像中画幅占比较小的问题,使用马赛克增强的方式增加了数据集中小目标杆塔的数量。其次,设计了基于深度学习的无人机巡检图像线路杆塔检测模型。该模型基于YOLOv3算法,并在原算法的基础上提出了对先验框和网络结构的改进,通过实验验证了算法改进的有效性,并比较了改进后的YOLOv3算法与YOLOv4算法的性能。最后,为了算法模型能够应用在计算和存储资源有限的嵌入式设备上,本文还利用通道剪枝技术对训练好的杆塔检测模型进行压缩,在损失少量精度的情况下,模型参数减少90%以上,进一步提高了检测速度。进一步,研究了基于无人机巡检图像的线路杆塔目标物理定位算法。结合图像元数据,利用坐标变换将无人机、线路杆塔像素点、线路杆塔实际空间位置,统一到空间直角坐标系中,并利用共线方程求解线路杆塔在空间直角坐标系中的坐标,然后将空间直角坐标系变换到大地坐标系,从而获得线路杆塔的GPS坐标。本文通过实例对该方法进行了评估,实验结果表明误差在合理范围内。另外,考虑到无人机元数据中存在传感器测量导致的随机误差,本文还利用蒙特卡洛方法验证了元数据存在误差的情况下,计算获得的杆塔GPS误差依然能够保持在合理区间,从而进一步证明了定位算法的有效性。最后,基于Python/Django框架设计并开发了无人机线路杆塔巡检图像识别与灾损信息统计软件系统。分别开发了图像数据管理、图像数据处理、灾损信息统计等算法模块和可视化操作界面,实现了端到端的灾后无人机线路杆塔巡检图像的目标检测与定位。更多还原

【Abstract】 Typhoons are frequent and destructive in the southern coastal areas of China,greatly affecting the safety and stability of the power grid.When the power grid is hit by typhoons,mudslides and other disasters,the towers of transmission and distribution lines are prone to tilting and collapsing,which can easily lead to line failures and consequently power outages.To quickly and accurately locate abnormal towers and restore power supply promptly,power operation and maintenance departments usually use unmanned aerial vehicles(UAVs)for post-disaster power tower inspections.However,the massive inspection images generated by UAV inspections are still mainly processed manually,which is not only time-consuming and laborious but also prone to omissions and misjudgments.Therefore,this dissertation adopts deep learning-based image recognition technology for intelligent detection of power poles and towers and uses the metadata of UAV inspection images to calculate the GPS information of power poles and towers through coordinate transformation,to provide auxiliary decision-making information for rapid repair of power poles and towers after disasters and improve line repair capability.The main research content of this dissertation is as follows.Firstly,the image enhancement technology applicable to UAV power pole and tower inspection images is studied,and a database of UAV power pole and tower inspection images of uniform category and moderate quantity is constructed.To improve the robustness of the model for image recognition under various environments,image enhancement methods such as distortion and filtering are used to extend the image samples according to the principle of VRM(Vicinal Risk Minimization).To solve the problem of the small size of power poles and towers in the whole inspection image,the number of small target towers in the data set is increased by using mosaic enhancement.Secondly,a deep learning-based power pole and tower detection model for UAV inspection images was designed.The model is based on the YOLOv3 algorithm and the improvements to the prior frame and network structure based on the original algorithm is proposed.The effectiveness of the algorithm improvements is verified through experiments and the performance of the improved YOLOv3 algorithm is compared with that of the YOLOv4 algorithm.Finally,in order to apply the algorithm model in embedded devices with limited computational and storage resources,this dissertation also uses channel pruning techniques to compress the trained pole tower detection model,reducing the model parameters by more than 90% with a small loss of accuracy,which further improves the detection speed.Thirdly,a power pole and tower physical localization algorithm based on UAV inspection images is studied.The algorithm firstly uses coordinate transformation based on the metadata of the image to unify the UAV,pixel point of the power pole or tower,and the actual physical position of the power pole or tower into the spatial rectangular coordinate system,then solves the collinear equation to obtain the coordinate of the power pole or tower in the spatial rectangular coordinate system,and finally transforms the coordinate to the geodetic coordinate system to obtain the GPS coordinates of the power pole or tower.In this dissertation,the method is evaluated by a case study and the experiments show that the error is within a reasonable range.Besides,considering the existence of random errors caused by sensor measurements in the UAV metadata,this dissertation also uses Monte Carlo methods to verify that the calculated GPS errors obtained for the pole towers remain within a reasonable range in the presence of errors in the metadata,further demonstrating the effectiveness of the method.Finally,a software system for the image recognition and disaster damage information statistics of UAV power pole and tower inspection was designed and developed based on the Python/Django framework.The algorithm modules and user interfaces for image data management,image data processing,and disaster damage information statistics were designed respectively to achieve end-to-end target detection and location of UAV power pole and tower inspection images after disasters.更多还原