【作者】 吴迪；

【导师】 李辰；

【作者基本信息】 哈尔滨工业大学 ， 电子信息（专业学位）， 2022， 硕士

【摘要】 严重覆冰会引起线路舞动、绝缘子闪络、杆塔倾斜甚至坍塌等事故灾害,威胁输电线路系统安全运行时,可能会对生产生活造成巨大冲击。据此,基于覆冰厚度检测准确获得厚度信息,并结合其它算法获得覆冰类别信息,为电力人员判断覆冰状态提供可靠依据,保障电力系统安全运行。目前,覆冰厚度检测及类别研究大多依赖人工巡检或传感器装置,受环境因素影响较大且损耗率高。根据调研,我国南方电网与国家电网管辖地区均设置了覆冰监测系统,以监测系统获得的覆冰图像为研究对象,针对传统图像处理算法泛化能力及鲁棒性欠佳等问题,本文提出基于深度学习实现覆冰分割与分类,在分割基础上实现覆冰等效厚度计算并完成软件功能开发,对实现电网智能化具有重要意义。本文围绕覆冰厚度检测与分类研究目标,分别提出了覆冰分割网络模型与分类网络模型,并提出了在分割基础上的覆冰等效厚度计算方案。首先收集原始覆冰图像,对其进行预处理操作后搭建覆冰图像数据集。针对覆冰分割模型,基于U-Net编码解码结构,从主干网络选取、卷积结构轻量化、注意力机制、迁移学习及辅助损失等方面进行改进,提出覆冰分割网络S＿UNet。通过实验分析,S＿UNet误检、漏检现象仍有待改进;将生成对抗网络(Generative Adversarial Networks,GAN)与S＿UNet融合提出覆冰分割优化网络SGAN＿UNet,利用真实标签图对覆冰图像分割结果进行比对校正,提升了覆冰分割结果精细程度。针对覆冰分类,基于Efficient Net B3网络,结合数据增强及标签平滑技术搭建输电线路覆冰分类网络,完成对雨凇、混合淞、雾凇、裸线路与无线路五种类别的分类。针对覆冰等效厚度计算,以同一角度拍摄的覆冰前后分割图像为研究对象,建立了覆冰厚度计算模型,考虑覆冰图像或存在透视畸变,提出基于逆透视变换(Inverse Perspective Mapping,IPM)对畸变进行校正,并从原理及分割误差角度对覆冰厚度计算进行了实验验证。最后开发了具有覆冰厚度检测与分类功能的软件系统。更多还原

【Abstract】 Severe ice-covered disasters can lead to line galloping,insulator flashing,tower tilting and even collapse accidents,it threatens the safe operation of the transmission line system and has a serious impact on production and life.Accordingly,the ice thickness information can be accurately obtained based on the ice-cover detection,combined with the ice-cover classification information,to provide a credible basis for electric personnel to judge the ice-covered state and guarantee the secure operation of the power system.At present,most of the ice-cover thickness detection and classification studies rely on manual inspection or sensor devices,which are greatly affected by environmental factors and have a high loss rate.According to the research,both the Southern Power Grid and the National Power Grid have established ice-cover monitoring systems.The icecover image information obtained from the monitoring system is used as the research object,to address the problems of penurious generalization ability and robustness of traditional image processing algorithms,this paper proposed to realize the segmentation and classification of ice-cover based on deep learning,and to calculate the equivalent thickness of ice-cover on the basis of segmentation and realize software function development,which is important for achieving grid intelligence.In this paper,focusing on the research objectives of ice-covered thickness detection and classification,the ice-covered segmentation network model and classification network model are proposed respectively,and the equivalent thickness calculation scheme of ice-covered on the basis of segmentation is proposed.Firstly,after collecting the original ice-covered images,pre-processing operations are performed to build the ice-covered image dataset.Based on the UNet coding and decoding structure,this paper proposes the ice-covered segmentation network S＿UNet by improving the selection of the backbone network,convolutional structure lightweighting,attention mechanism,transfer learning and auxiliary loss,etc.Through the experimental analysis,the phenomenon of S＿UNet misdetection and omission still needs to be improved.In this paper,generative adversarial network(GAN)and S＿UNet are combined to propose an ice-covered segmentation optimization network SGAN＿UNet,which uses the real label map to compare and correct the ice-covered image segmentation results.The fineness of the ice-covered segmentation results is improved.For icecover classification,this paper builds a transmission line ice-cover classification network based on Efficient Net B3 network,combined with data enhancement and label smoothing technology,and completes the classification of five categories of rain,mixed,smog,bare lines and no lines.In order to calculate the equivalent thickness of ice cover,this paper establishes a model for calculating the thickness of ice cover based on the front and back segmentation images taken from the same angle.Considering that there may be perspective aberrations in ice cover images,this paper propose to correct the aberrations based on Inverse Perspective Mapping(IPM),and experimentally verify the ice cover thickness calculation from the perspective of principle and segmentation error.Finally,a software system with ice cover thickness detection and classification functions was developed.更多还原