基于图像处理的接触网绝缘子裂纹和定位支座检测

发布时间：2018-04-26 06:22

本文选题：绝缘子 + LBP　；参考：《西南交通大学》2017年硕士论文

【摘要】：接触网及其附属部件的良好工作状态是高速列车安全运行的基本保障。由于其长时间处于工作状态,接触网的绝缘子和定位支座不断受到电气冲击及机械应力的影响,如果绝缘子和定位支座发生故障,轻则损坏接触网设备,重则将会直接导致高速列车的骤停甚至造成人员伤亡。目前,基于图像处理技术的绝缘子工作状态检测已有研究,这大大提高了设备巡检效率,节省了不少人力财力。然而,以往的检测方法基本是针对绝缘子片缺失或夹杂异物的研究,绝缘子破裂故障的研究较少,且定位支座的检测几乎为空白,因此有必要采用图像处理中机器学习的高效方法检测绝缘子和定位支座。在对以往接触网及附属装置检测方法、电力系统中绝缘子的识别进行研究以及对比人脸识别方法之后,本文利用接触网综合巡检车采集的图像数据作为样本的原始图像,利用LBP和HOG提取绝缘子的局部特征,接着采用机器学习的方法训练分类器对图像中的绝缘子进行精确提取,然后对绝缘子裂纹进行分析。同时采用基于ASIFT、SURF、ORB、FREAK四种特征匹配的方法实现了小目标定位支座的检测。首先对绝缘子的原始图像做了形态学运算等相关预处理,建立了以大量绝缘子目标和非绝缘子图像为基础的正负样本库,然后提取绝缘子的LBP和HOG特征。利用机器学习的方法提取图像中的目标绝缘子。分别将提取的LBP、HOG绝缘子的特征交给Opencv利用Adaboost算法训练出分类器,然后利用分类器模型在图像中进行绝缘子定位识别,对比发现LBP与Adaboost组合模型的绝缘子识别率最高。最后运用此模型对大量接触网图像进行绝缘子精确提取,利用多种经典边缘检测的方法和Canny检测算子提取出目标绝缘子边缘,对比发现Canny算子的检测效果最好,利用阈值化方法对角度校正后的绝缘子进行二值化,分割绝缘子和裂纹,采用连通域求面积和周长的方法计算绝缘子裂纹的几何特征,从而实现绝缘子裂纹检测。在定位支座的识别中,运用多种特征匹配的算法对定位支座进行定位识别,发现SURF算子性能更好。实验是利用OpenCV2.4.13库以及软件VS2013编程,通过对大量接触网图像进行实验测试,得出了 LBP与Adaboost模型在绝缘子检测时的有效性、绝缘子裂纹分析的准确性以及SURF算子对定位支座检测的可靠性。
[Abstract]:The good working state of the contact network and its accessory parts is the basic guarantee for the safe operation of the high-speed train. Because of its long working condition, the insulator and the positioning support of the contact network are constantly affected by the electrical shock and mechanical stress. If the insulators and the positioning support fail, the contact network equipment will be damaged lightly and the weight will be straight. The sudden stop of high-speed train even causes casualties. At present, the detection of the working state of Insulators Based on image processing technology has been studied, which greatly improves the efficiency of the equipment inspection and saves a lot of manpower and financial resources. However, the previous detection methods are mainly aimed at the lack of insulators or the inclusion of foreign objects, and the rupture of insulators. There are few studies on the obstacle, and the detection of the location support is almost blank, so it is necessary to use the efficient method of machine learning in the image processing to detect the insulators and the positioning support. The image data collected by the comprehensive toured patrol vehicle is used as the original image of the sample. Using LBP and HOG to extract the local characteristics of the insulators, the classifier is trained by machine learning to extract the insulators in the image accurately, and then the insulator cracks are analyzed. Four features based on ASIFT, SURF, ORB, FREAK are used in the same time. The matching method realizes the detection of small target location support. First, the original image of insulators is preprocessed by morphological operation. A positive and negative sample library based on a large number of insulators target and non insulator image is established. Then the LBP and HOG features of insulators are extracted and the object in the image is extracted with machine learning method. Insulators. The features of the extracted LBP and HOG insulators are given to Opencv to train the classifier using the Adaboost algorithm. Then the classifier model is used to identify the insulators in the image. It is found that the insulator recognition rate of the LBP and Adaboost combination model is the highest. Finally, the model is used to insulators for a large number of contact network images. Accurate extraction, using a variety of classical edge detection methods and Canny detection operators to extract the edge of the target insulator, the contrast found that the detection effect of the Canny operator is the best. Using the threshold method, the insulator and the crack are divided into two values, the insulators and the cracks are segmented, and the area and the circumference of the connected domain are used to calculate the insulators. The geometric characteristics of the crack can be used to detect the insulator crack. In the identification of the positioning support, a variety of feature matching algorithms are used to locate the positioning support, and the SURF operator is found to be better. The experiment is to use the OpenCV2.4.13 library and software VS2013 to test the large amount of catenary images and get the LBP The validity of the Adaboost model in insulator detection, the accuracy of insulator crack analysis, and the reliability of SURF operator for locating bearing detection are also discussed.

【学位授予单位】：西南交通大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：U226.8;TP391.41

【参考文献】