输电线路上鸟巢的检测算法研究
发布时间：2019-02-23 15:56
【摘要】：输电线路上的鸟巢检测是智能电网中智能巡检的重要研究内容,鸟类在输电线路杆塔上的筑巢会对输电线路等设备造成不良影响,甚至危害电网的安全运行。然而传统的人工巡检输电线路的方式耗时耗力且存在危险,给电力工作者带来很大的困扰。因此,急需对输电线路的自动检测技术和系统。本文在调研国内外相关工作基础上,设计了基于计算机视觉和机器学习的输电线路鸟巢图像自动检测系统,为电网智能巡检提供算法和技术。本文首先根据已有数据集建立两千多张含有鸟巢的高压输电线路图像数据库,并将其从整体上分为两大类:(1)简单图像:鸟巢枝条明显裸露的图像;(2)复杂图像:鸟巢枝条模糊不清。针对这两大类图像,本文分别设计了两类不同的解决方案,并进行了实验对比和分析。最终有监督算法中基于深度学习的方案可以获得更高的检测性能。本文的主要工作有:(1)提出了基于K-Means以及GMM(Gaussian Mixture Model)的无监督鸟巢检测算法,并在第一类图像上验证。对于第一类图像,首先进行预处理,即去除干扰物,留下鸟巢枝条部分。然后利用渐进式霍夫变换提取线条,针对鸟巢这一类特定目标,设计了鸟巢枝条长度直方图与方向直方图特征,最终结合PCA(Principal Components Analysis)实现无监督鸟巢识别的目的。无监督方案不需要进行大量样本的机械化标注工作,也不需要进行复杂的训练过程,且算法实现效率高,但因其鲁棒性不强,本文还设计了三种有监督的方案进行鸟巢检测,并在两类图像集上进行定量验证。(2)对从第一类图像中收集到的直方图信息进行人工分类并加注标签,然后将其作为学习样本输入到KNN(K-NearestNeighbor)算法中进行训练,训练完成后对未知标签样本进行两种参数下的实验对比;(3)从两类图像中截取鸟巢样本以及非鸟巢样本,基于两类样本的Haar特征及LBP(Local Binary Pattern)特征训练AdaBoost分类器,实验表明基于LBP特征的分类器可以达到更高的检测准确率;(4)利用两类数据集对FastR-CNN深度学习中的CaffeNet网络进行微调,进而训练出适用于检测鸟巢的神经网络模型。通过实验验证,该方法可以获得92.46%的检测准确率。通过实验结果的对比与分析,本文最终选择有监督深度学习方案。因为相比于其他方案,此方案具有更强的适用性,且对于鸟巢枝干的遮掩以及鸟巢形状、大小、光线强弱等都有一定的鲁棒性。
[Abstract]:Bird nest detection on transmission line is an important research content of intelligent inspection in smart grid. Nest building on transmission line tower will cause adverse effects on transmission line and even endanger the safe operation of power grid. However, the traditional manual inspection of transmission lines is time-consuming, time-consuming and dangerous, which brings a lot of trouble to power workers. Therefore, the automatic detection technology and system of transmission line are urgently needed. On the basis of investigation and research at home and abroad, this paper designs an automatic detection system for bird's nest image of transmission lines based on computer vision and machine learning, which provides algorithm and technology for intelligent inspection of power grid. In this paper, more than two thousand HV transmission line images with bird nests are established according to the existing data sets, and they are divided into two categories: (1) simple images: the obvious naked images of the branches of bird's nests; (2) complex image: the branches of the nest are blurred. For these two kinds of images, this paper designs two kinds of different solutions, and carries on the experiment contrast and the analysis. Finally, the supervised algorithm based on depth learning can achieve higher detection performance. The main work of this paper is as follows: (1) an unsupervised nest detection algorithm based on K-Means and GMM (Gaussian Mixture Model) is proposed and verified on the first kind of images. For the first kind of image, the first image is preprocessed, that is, the interference is removed and the nest branch is left. Then the progressive Hough transform is used to extract the lines and the histogram of the length of the branch and the direction histogram of the bird's nest are designed for the specific target of the bird's nest. Finally the purpose of unsupervised nest recognition is realized with PCA (Principal Components Analysis). The unsupervised scheme does not need a large number of samples to carry out mechanization marking work, and does not need to carry on the complex training process, and the algorithm realization efficiency is high, but because its robustness is not strong, this paper also designs three kinds of supervised schemes to carry on the bird nest detection. The histogram information collected from the first kind of images is classified manually and tagged, and then input into the KNN (K-NearestNeighbor) algorithm as a learning sample for training. After the completion of the training, the unknown label samples were compared with each other under two kinds of parameters. (3) intercepting bird nest samples and non-nest samples from two kinds of images, training AdaBoost classifier based on Haar feature and LBP (Local Binary Pattern) feature of two kinds of samples, the experiment shows that the classifier based on LBP feature can achieve higher detection accuracy; (4) two kinds of data sets are used to fine-tune the CaffeNet network in FastR-CNN depth learning, and then a neural network model suitable for detecting bird's nest is trained. Experimental results show that this method can obtain 92.46% accuracy. Through the comparison and analysis of the experimental results, this paper finally chooses the supervised deep learning scheme. Compared with other schemes, this scheme is more applicable and robust to the shelter of the nest branches, the shape, size and light intensity of the nest.
【学位授予单位】：北京交通大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TM755
本文编号：2428975
[Abstract]:Bird nest detection on transmission line is an important research content of intelligent inspection in smart grid. Nest building on transmission line tower will cause adverse effects on transmission line and even endanger the safe operation of power grid. However, the traditional manual inspection of transmission lines is time-consuming, time-consuming and dangerous, which brings a lot of trouble to power workers. Therefore, the automatic detection technology and system of transmission line are urgently needed. On the basis of investigation and research at home and abroad, this paper designs an automatic detection system for bird's nest image of transmission lines based on computer vision and machine learning, which provides algorithm and technology for intelligent inspection of power grid. In this paper, more than two thousand HV transmission line images with bird nests are established according to the existing data sets, and they are divided into two categories: (1) simple images: the obvious naked images of the branches of bird's nests; (2) complex image: the branches of the nest are blurred. For these two kinds of images, this paper designs two kinds of different solutions, and carries on the experiment contrast and the analysis. Finally, the supervised algorithm based on depth learning can achieve higher detection performance. The main work of this paper is as follows: (1) an unsupervised nest detection algorithm based on K-Means and GMM (Gaussian Mixture Model) is proposed and verified on the first kind of images. For the first kind of image, the first image is preprocessed, that is, the interference is removed and the nest branch is left. Then the progressive Hough transform is used to extract the lines and the histogram of the length of the branch and the direction histogram of the bird's nest are designed for the specific target of the bird's nest. Finally the purpose of unsupervised nest recognition is realized with PCA (Principal Components Analysis). The unsupervised scheme does not need a large number of samples to carry out mechanization marking work, and does not need to carry on the complex training process, and the algorithm realization efficiency is high, but because its robustness is not strong, this paper also designs three kinds of supervised schemes to carry on the bird nest detection. The histogram information collected from the first kind of images is classified manually and tagged, and then input into the KNN (K-NearestNeighbor) algorithm as a learning sample for training. After the completion of the training, the unknown label samples were compared with each other under two kinds of parameters. (3) intercepting bird nest samples and non-nest samples from two kinds of images, training AdaBoost classifier based on Haar feature and LBP (Local Binary Pattern) feature of two kinds of samples, the experiment shows that the classifier based on LBP feature can achieve higher detection accuracy; (4) two kinds of data sets are used to fine-tune the CaffeNet network in FastR-CNN depth learning, and then a neural network model suitable for detecting bird's nest is trained. Experimental results show that this method can obtain 92.46% accuracy. Through the comparison and analysis of the experimental results, this paper finally chooses the supervised deep learning scheme. Compared with other schemes, this scheme is more applicable and robust to the shelter of the nest branches, the shape, size and light intensity of the nest.
【学位授予单位】：北京交通大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TM755
【参考文献】
相关期刊论文 前4条
1 段旺旺;唐鹏;金炜东;韦璞;;基于关键区域HOG特征的铁路接触网鸟巢检测[J];中国铁路;2015年08期
2 杨振伟;代晓光;谢平;;九江地区输电线路鸟害故障分析与防治[J];华中电力;2011年03期
3 王少华;叶自强;;架空输电线路鸟害故障及其防治技术措施[J];高压电器;2011年02期
4 郭伟跃;;美国输电线路和变电站电气设备防鸟害措施[J];中国电力;2006年08期
，本文编号：2428975
本文链接：https://www.wllwen.com/kejilunwen/dianlidianqilunwen/2428975.html
