【作者】 陈实；

【导师】 马草原；

【作者基本信息】 中国矿业大学 ， 电气工程（专业学位）， 2024， 硕士

【摘要】 随着我国能源转型不断推进和电力需求的日益增长,电力输送领域对高效可靠巡检技术的需求日益紧迫。绝缘子在航拍图像中存在环境复杂多样、缺陷目标较小等问题,给绝缘子及其缺陷的检测带来了极大挑战。本文基于改进的YOLOv8算法在提高模型精度和模型轻量化两方面进行研究,最终成功训练出了两套针对绝缘子缺陷的检测模型,并基于Py Side2成功将改进算法嵌入绝缘子检测系统。具体工作内容如下:(1)针对现有绝缘子数据集存在噪声干扰、图像质量不足以及样本数量不均衡等问题,采用降噪、数据增强和样本扩增等技术,并通过Label Img工具完成数据标注,实现高质量数据集的制作。(2)对基于YOLOv8的输电线路绝缘子缺陷检测算法进行性能优化研究。针对绝缘子缺陷普遍较小、表观特征较难识别的问题,提出在网络结构中加入Bi Former注意力机制,使模型能够过滤掉特征图中大多数低相关区域,将注意力集中在高相关特征上。针对复杂背景下难以有效捕获绝缘子缺陷相关特征的问题,引入可变形Conv Nets v2模块来增强算法对几何空间变换的适应性,并将LSKNet模块嵌入到深度网络结构,采用加权融合的方式生成特征融合网络CNL,使模型能够更好地适应不同的背景信息。针对YOLOv8的原损失函数CIo U收敛过程波动较大的问题,选择更为稳健的损失函数WIo U v3,减少极端样本中较大或有害梯度的发生,提高模型的泛化能力。通过对改进方法的组合应用,得到改进的BL-YOLOv8模型。实验结果表明BL-YOLOv8模型m AP0.5/%提高了2%,m AP0.5:0.95/%提高了6.6%,验证了其优越性。(3)针对平台资源有限的问题,采用深度可分离卷积、通道混洗以及Faster Net模块重新搭建YOLOv8主干特征网络,通过实验对比最后选用Faster Net中的残差模块Faster Net Block替换C2f模块中的Bottle Neck,并用一种小目标检测层去替换大目标检测层,成功对YOLOv8模型进行轻量化改进。实验结果表明轻量化模型F-YOLOv8较基础模型参数量下降了51.2%,计算量下降了35.5%,能更有效地部署在移动设备上。(4)设计了一个应用改进YOLOv8算法的绝缘子检测系统。针对传统人工管理大数据耗时长,且后期查询巡检数据不便等问题,设计开发了基于Py Side2的绝缘子缺陷检测系统软件,主要包含了登录、模型更新、检测、系统日志四大模块。经过功能测试,该系统符合设计要求,可以满足绝缘子缺陷检测的需求。更多还原

【Abstract】 As Chinese energy transition continues to advance and the demand for electricity grows,the demand for efficient and reliable inspection technology in the field of power transmission is becoming increasingly urgent.Insulators in aerial images have problems such as complex and diverse environments and small defect targets,which bring great challenges to the detection of insulators and their defects.In this thesis,based on the improved YOLOv8 algorithm to improve the model accuracy and model lightweight two aspects of research,and finally successfully trained two sets of insulator defect detection model for insulators,and based on Py Side2 successfully embedded the improved algorithm into the insulator detection system.The specific work is as follows:(1)Aiming at the existing insulator dataset’s problems such as noise interference,insufficient image quality and uneven number of samples,the production of highquality dataset is achieved by adopting the techniques of noise reduction,data enhancement and sample augmentation,and completing the data annotation through the Label Img tool.(2)Research on performance optimisation of insulator defect detection algorithm for transmission lines based on YOLOv8.For the problem that insulator defects are generally small and the apparent features are difficult to identify,it is proposed to add the Bi Former attention mechanism to the network structure,so that the model can filter out most of the low correlation regions in the feature map,and focus the attention on the high correlation features.Aiming at the problem that it is difficult to effectively capture insulator defect-related features under complex backgrounds,the deformable Conv Nets v2 module is introduced to enhance the adaptability of the algorithm to geometric space transformations,and the LSKNet module is embedded into the deep network structure,and the weighted fusion is used to generate the ad-hoc fusion network CNL,so that the model can better adapt to different background information.Aiming at the problem of large fluctuations in the convergence process of the original loss function CIo U of YOLOv8,a more robust loss function WIo U v3 is chosen to reduce the occurrence of large or harmful gradients in extreme samples and to improve the model’s generalisation ability.Through the combined application of the improved methods,the improved BL-YOLOv8 model is obtained.The experimental results show that the BL-YOLOv8 model m AP0.5/% improves by 2% and m AP0.5:0.95/% improves by 6.6%,which verifies its superiority.(3)Aiming at the problem of limited platform resources,the YOLOv8 backbone feature network is reconstructed by using deep separable convolution,channel mixing and Faster Net module,and finally the residual module Faster Net Block in Faster Net is selected to replace Bottle Neck in C2 f module through experimental comparison,and a small target detection layer is used to Replacing the large target detection layer with a small target detection layer,the YOLOv8 model is successfully lightweighted and improved.The experimental results show that the lightweight model F-YOLOv8 reduces the number of parameters by 51.2% and the computation amount by 35.5%compared with the baseline model,and can be deployed on mobile devices more effectively.(4)An insulator detection system applying the improved YOLOv8 algorithm was designed.Aiming at the problems of time-consuming traditional manual management of big data and inconvenience of querying inspection data at a later stage,the insulator defect detection system software based on Py Side2 is designed and developed,which mainly contains four major modules,namely,login,model updating,detection,and system log.After functional testing,the system meets the design requirements and can meet the needs of insulator defect detection.更多还原