【作者】 李天伟；

【导师】 陈中； 马艳；

【作者基本信息】 东南大学 ， 电气工程（专业学位）， 2021， 硕士

【摘要】 随着电力行业规模、形态不断发展进步,依靠电力系统变电站值班员开展的人工巡视逐渐无法满足变电站日常运行维护的要求,近年来,特别是“十三五”期间,国内经济社会发展水平持续提高,全社会用电负荷及变电站数量的快速增长,给变电站设备运检工作带来了新的要求和挑战。以智能机器人巡检为主要模式的机器代人技术是电力行业发展的必然趋势。图像识别及处理技术是变电站智能巡检机器人的关键技术之一,目前应用在变电站巡检领域的图像识别技术尚未完全成熟,图像识别效果仍有一定的进步空间。本文重点研究了高压断路器分合指示、避雷器泄漏电流表读数、避雷器动作次数数字三种变电站状态量识别过程,通过对不同机器学习算法的识别结果详细比较,得出了针对某种电力设备状态的最佳识别算法,此研究成果可以应用在智能巡检机器人的图像识别系统中,提高了智能巡检机器人的巡检质量,减少误识别、低质量识别概率,从而提高效率,降低成本。论文主要内容包括:(1)阐述了变电站智能巡检中图像识别与处理技术的研究背景,结合电网公司实际状况详细分析了以巡检机器人为代表的智能巡检技术将逐步取代变电站人工巡检。分析了智能巡检图像识别领域国内外研究现状,总结出当前应用在电力系统巡检机器人系统中的图像识别方法主要存在的几个方面的不足,在变电站指针式仪表读数、文字及数字智能识别方面还有许多新的难题需要解决。(2)以华东某地市220千伏东枫变电站应用的智能机器人为例,介绍了变电站智能巡检系统组建方案。巡检机器人检测表盘时最常用的方法是模板匹配法,模板匹配法可以在一幅整体图像中寻找某个特定目标,模板匹配法的原理相对简单,即计算图片中各处与模板的“相似度”,“相似度”越高,则越接近要找的目标。详细统计了220千伏东枫变电站内智能巡检机器人针对高压断路器分合指示、避雷器泄漏电流表读数、避雷器动作次数数字三种变电站状态量的实际识别效果,由统计结果可知,在识别过程中,除外图像拍摄失败等因素,存在一定数量的识别不成功的情况,导致在智能巡检机器人现场实际应用过程中,运维人员需对疑似识别失败的状态量开展现场复核。表明应用在220千伏东枫变电站的智能巡检机器人对于上述三种状态量的识别效果还有一定的提升空间。(3)研究了变电站高压断路器分合状态检测方法。高压断路器是电力系统变电站中主要的电气设备之一,高压断路器的安全可靠运行对于维护电力系统稳定具有重要意义。针对采集到的断路器分合状态的图像,对图像进行标记、分类,支持向量机(support vector machine,SVM)算法可以自行找到“分”与“合”分类超平面关系密切的众多支持向量,最大化类间隔,提高分类能力,具有较好的分辨力。通过与基于朴素贝叶斯算法的检测模型比较,基于支持向量机的开关分合指示状态检测模型具有较高的检测率。(4)研究了避雷器泄漏电流指针读数状态检测方法。电力系统变电站中各类指针式表计是巡视、设备监测的重要设备,主要有避雷器泄漏电流表、设备SF6压力表、设备油位表、变压器油温度表、变压器绕组温度表、电压表、电流表等,针对指针式表计的巡视任务占据变电站巡视的大部分工作量,也是智能机器人巡检的重要内容。针对避雷器泄漏电流仪表盘包含状态量较多的情况,利用SURF目标检测技术在原始照片中检测出完整的子仪表盘图像,实现泄漏电流子表盘图像的检测与抓取,便于后续泄漏电流读数识别。同时,利用k-最近邻(KNN)算法,将训练集作为KNN模型的输入,计算分类目标与样本点之间的距离,统计距离最近的前K个样本点所属各类别占比,选择占比最大的类别作为类别结果。实验证明,相对于基于SURF-SVM读数状态检测模型,基于SURF-KNN的检测模型检测准确率较高。(5)提出一种基于SURF-随机森林的避雷器动作次数数字检测方法。避雷器的动作次数可以反映电力设备遭受雷击的次数,可以指导运维人员针对性开展设备运维,为线路运行维护提供参考。利用SURF目标检测技术在原始照片中检测出完整的动作次数显示器图像;然后利用随机森林算法识别数字对象,将训练集作为随机森林算法的输入,利用多数投票原理,训练出避雷器动作次数数字检测模型,然后将测试集作为模型的输入,对训练好的模型进行测试。实验证明,利用SURF-随机森林检测方法良好的目标检测和识别能力,并对比随机森林算法与决策树算法识别的准确率,验证了识别方法的准确度与稳定度。更多还原

【Abstract】 The main content of this subject is the research and application of intelligent inspection image recognition technology for substations.Modern society is increasingly dependent on electricity,and the stable and safe operation of the power system is essential.With the continuous development of the power industry,manual inspections have become increasingly unable to meet the requirements of safe operation of modern substations.The replacement of manual inspections by intelligent inspection robots will be an inevitable trend in the development of power grid technology.The key technology of intelligent inspection robots is image recognition and processing.Technology,the image recognition technology currently applied in the field of substation inspection still has certain limitations and research gaps.Aiming at the key equipment in the operation of the substation,this paper systematically studies the image recognition method of the opening and closing status of the circuit breaker,the reading of the arrester instrument,and the number of arresters.The main content of this article includes:(1)The background and significance of the research on intelligent inspection image recognition technology for substation equipment are described,and the current research status and existing problems of intelligent inspection image recognition technology at home and abroad are analyzed.The composition of the intelligent robot inspection system and the actual effect of the robot inspection image recognition in the 220 k V Dongfeng substation in East China are introduced.(2)The detection method of switching status of high-voltage circuit breakers in substations is studied.According to the collected switch state images,the images are marked and classified,and the substation intelligent inspection switch state model of the Support Vector Machine is trained to verify the accuracy and robustness of the model.Compare the recognition accuracy of the support vector machine algorithm and Naive Bayes algorithm.(3)The detection method of the reading status of the arrester leakage current pointer is studied.The Speeded-Up Robust Features algorithm is used to detect and capture the arrester leakage current sub-dial image to ensure that the complete arrester leakage current sub-dial image can be accurately collected.The k-Nearest Neighbor algorithm is used to identify the leakage current readings of the arrester,to verify that the meter reading recognition method has high accuracy and stability,and to compare the accuracy of the k-Nearest Neighbor algorithm and the Support Vector Machine algorithm.(4)The digital detection method for the number of arrester actions is studied.Firstly,use the Speeded-Up Robust Features algorithm to detect and capture the sub-images of the lightning arrester leakage current meter action times display,ensuring that the complete lightning arrester leakage current meter action times display sub-images can be accurately collected.The Random Forest algorithm is used to identify the number of lightning arrester actions,to verify that the recognition method has high accuracy and stability,and to compare the accuracy of the Random Forest algorithm and the Decision Tree algorithm.更多还原