【作者】 史浩；

【导师】 聂礼强；

【作者基本信息】 山东大学 ， 计算机技术（专业学位）， 2021， 硕士

【摘要】 随着十四五规划的展开,稳定的电力供应成为保障经济社会正常运转的关键一环。销钉是输电线路中用于固定螺母的器件,销钉的脱落会导致输电线路的不稳定,极易引起跳闸事故。近几年,基于深度神经网络的目标检测技术获得了飞速发展,尤其在电力运维中与无人机巡检进行结合,提高了巡检人员的巡检效率和人身安全性。因此一种基于深度神经网络的销钉缺陷检测方法对巡检人员完成销钉缺陷的巡检工作,对维护输电安全具有重要的研究意义和应用价值。销钉缺陷检测主要存在三个挑战,最大的挑战在于销钉是绝对意义上的小物体;其次,复杂的自然环境背景和邻近的大量相似机械部件加剧了检测难度;第三,在数据层面上存在正常状态销钉和缺失状态销钉的类别不平衡问题。当前的销钉缺陷检测研究分为基于传统图像处理技术的检测方法和基于深度神经网络的目标检测模型的方法,但检测销钉缺陷的性能依然不能满足工业界的实际需求。本文介绍一种基于多层级联结构的销钉缺陷检测方法,其在性能表现上优于直接使用单个深度神经网络的目标检测模型进行销钉缺陷检测的方法。本文的检测方法分为四个模块,分别为图像预处理模块、冗余滑动窗口分割模块、销钉图像定位模块和销钉状态分类模块。通过多层级联的方式,不仅逐步提高销钉图像在待检测图像中的区域占比,放大销钉图像的特征,还能够过滤掉无关的复杂背景。为训练和测试本文提出的销钉缺陷检测方法,本文构建了三个基于真实输电线路场景下的销钉样本数据集,在构建过程中采用数据增强的方式对销钉样本数据集中的类别不平衡的问题进行了改进。本文提出的基于多层级联结构的销钉缺陷检测方法和其中的重要模块在数据集上进行了大量实验,实验结果表明,本方法在检测缺失销钉上具有很高的准确率,并且针对改进部分的消融实验验证了其有效性。此外,本文面向巡检人员设计和实现了一个销钉缺陷检测系统。系统使用Python语言进行开发,采用B/S架构,前后端分离的开发方式。前端选用Vue等技术实现上传巡检图、调用部署的销钉缺陷检测方法以及查看对销钉的检测结果等界面。后端选用Flask作为Web框架,使用关系型数据库MySQL保存巡检图和检测结果,使用gRPC的方式调用通过Paddle Serving部署的销钉缺陷检测模型。更多还原

【Abstract】 With the launch of the 14th Five-Year Plan,a stable power supply has become a key part of ensuring the normal operation of the economy and society.The pin is a device used to fix the nut in the transmission line.The fall of the pin will cause the instability of the transmission line and easily cause a trip accident.In recent years,the object detection method based on the deep neural network has been developed rapidly,especially in the power operation and maintenance combined with UAV inspection,which improves the inspection efficiency and personal safety of inspectors.Therefore,a pin defect detection method based on the deep neural network has important research significance and application value for the inspection personnel to complete the pin defect inspection work,and to maintain power transmission safety.There are three main challenges in the pin defect detection.The biggest challenge is that pins are small objects in an absolute sense;secondly,the complex natural environment background and a large number of similar mechanical parts in the vicinity increase the difficulty of detection;thirdly,there exists the class imbalance problem between the normal status pins and missing status pins.Current research on the pin defect detection is divided into detection methods based on traditional image processing technology and the object detection model methods based on deep neural networks,but the performance of detecting pin defects still cannot meet the actual needs of the industry.In this paper,a method of pin defect detection based on multi-level cascading structure is introduced.Its performance is better than that of using single deep neural network object detection model directly to detect pin defect.The detection method in this paper is divided into four modules,which are image preprocessing module,redundant sliding window segmentation module,pin image positioning module and pin state classification module.By means of the multi-level cascading,not only the proportion of the pin image in the image to be detected is gradually increased,the features of the pin image are magnified,but also the unrelated complex background can be filtered out.In order to train and test the pin defect detection method proposed in this paper,three pin sample data sets based on real transmission line scenarios are constructed.In the construction process,the problem of class imbalance in the pin sample data set is improved by using the method of data enhancement.A large number of experiments have been carried out on the data set of the proposed method and its important modules.The experimental results show that the proposed method has a high accuracy in the detection of missing pins,and its effectiveness has been verified in the ablation experiment of the improved part.In addition,this paper designs and implements a pin defect detection system for inspection personnel.The system uses Python language to develop,using B/S architecture,front and rear end separation of development.Vue and other technologies are used in the front end to realize the interface of uploading the inspection map,calling the deployed pin defect detection method and viewing the test results of pins.At the back end,Flask was selected as the Web framework,the relational database MySQL was used to store the patrol maps and inspection results,and the pin defect detection model deployed via Paddle ad-serving was invoked using the gRPC.更多还原