基于卷积神经网络的电力巡检绝缘子检测研究

发布时间：2018-07-16 08:19

【摘要】：电力巡检是保障电网安全运行不可或缺的手段,新兴的无人机巡检通过搭载的高清相机和图传设备可获取大量详实的巡检影像。这些巡视数据仅凭人工分析和处理,工作量庞大,效率低下,存在由工作人员经验和素质引起的偏差。而绝缘子是电力系统中的常见部件,由于常年暴露在外,因而故障多发,严重威胁电网安全,需要引入智能化的识别方法自动进行故障诊断。本文结合四川省电力公司科技项目的需求,从以下几个方面展开研究:(1)本文通过搭建和改进卷积神经网络实现对绝缘子的检测,解决传统检测算法鲁棒性差,泛化能力不强,准确率不高等问题。首先通过研究卷积神经网络的特点和广泛应用,结合工程需求和硬件支持,完成对卷积神经网络各个部件的选型和设计,搭建适宜本课题网络模型。其次利用无人机在不同线路和时间采集玻璃和陶瓷绝缘子样本并进行人为拓展,作为训练样本。然后本文选择开源的Caffe作为工具,结合相关调参技术对网络结构进行改进和在训练过程中进行优化。通过自动学习绝缘子特征的本质和分布式表达,实现在复杂航拍背景中的绝缘子检测,训练准确率为95%,测试准确率为92%。(2)本文结合已训练完备的卷积神经网络完成绝缘子自爆的识别,解决人工分析工作量大,效率低等问题。首先利用卷积神经网络层级结构对全局和局部特征的综合与抽象,将训练完备的网络模型作为绝缘子特征抽取的工具,融入自组织特征映射网络,实现显著性检测的改良。其次在显著性检测的基础上,快速提取绝缘子,舍弃背景,然后结合超像素分割和轮廓检测等图像处理方法建立绝缘子模型,提出一种针对绝缘子自爆故障的识别算法,准确率在90%以上,取代人工分析,降低凭巡检工作人员经验判定的风险和误差,保障电网安全可靠运行。(3)本文对绝缘子检测及自爆故障识别分别进行测试验证和对比试验。首先针对不同背景,不同种类,不同数量的情况进行了绝缘子检测测试,并与传统的DPM和基于HoG的SVM算法进行对比。同时通过可视化效果分析网络的性能。然后对不同背景下的自爆识别算法进行了验证。最后以工程项目为依托,简单介绍电力巡检绝缘子检测系统平台的构架和应用效果。经验证,绝缘子检测和自爆识别均达到工程要求,有效体现巡视数据的价值,提升电力巡检的效率和智能化水平。
[Abstract]:Power inspection is an indispensable means to ensure the safe operation of the power grid. The emerging UAV patrol can obtain a large number of detailed inspection images by carrying high-definition cameras and graphic transmission equipment. These patrol data only depend on manual analysis and processing, the workload is huge, the efficiency is low, and the deviation caused by the staff's experience and quality exists. Insulator is a common component in power system. Because it is exposed all the year round, the fault is frequently occurred, which seriously threatens the security of power network, so it is necessary to introduce intelligent identification method to diagnose the fault automatically. According to the demand of Sichuan Electric Power Company's scientific and technological project, this paper studies the following aspects: (1) this paper realizes the detection of insulators by building and improving convolutional neural networks, which solves the problem of poor robustness and poor generalization ability of traditional detection algorithms. The accuracy is not high and so on. Firstly, by studying the characteristics and wide application of convolution neural network, combining with engineering demand and hardware support, the selection and design of each component of convolutional neural network are completed, and the network model suitable for this topic is built. Secondly, using UAV to collect glass and ceramic insulator samples on different lines and time, and to carry out artificial expansion, as a training sample. Then the open source Caffe is chosen as a tool to improve the network structure and optimize the training process. By automatically learning the nature and distributed expression of insulator features, the insulator detection in complex aerial photography background is realized. The accuracy of training is 95 and the accuracy of test is 92. (2) in this paper, the self-explosion identification of insulator is completed by using the fully trained convolution neural network to solve the problems of large workload and low efficiency in manual analysis. Firstly, using the hierarchical structure of convolution neural network to synthesize and abstract the global and local features, the well-trained network model is used as the tool of insulator feature extraction, and the self-organizing feature mapping network is integrated into the self-organizing feature mapping network to achieve the improvement of salience detection. Secondly, on the basis of salience detection, the insulator is quickly extracted and the background is discarded. Then the insulator model is established by combining the image processing methods such as super-pixel segmentation and contour detection, and an algorithm for identifying insulator self-detonation fault is proposed. The accuracy rate is over 90%, which replaces manual analysis, reduces the risk and error judged by the experience of inspection staff, and ensures the safe and reliable operation of power grid. (3) the insulator detection and fault identification of self-explosion are tested and compared in this paper. Firstly, the insulator detection test is carried out for different background, different kinds and different numbers of insulators, and compared with traditional DPM and SVM algorithm based on HoG. At the same time, the performance of the network is analyzed by visual effect. Then the self-detonation recognition algorithm under different background is verified. Finally, based on the project, the frame and application effect of the platform of insulator detection system for electric power inspection are briefly introduced. It is verified that both insulator detection and self-detonation identification meet the engineering requirements, effectively reflect the value of patrol data, and improve the efficiency and intelligence level of power inspection.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP183;TM755

【参考文献】