航拍绝缘子卷积神经网络检测及自爆识别研究

发布时间：2018-08-07 21:32

【摘要】：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。绝缘子是输电线路中极其重要且用量庞大的部件,在图像视频中快速准确地检测出绝缘子可为无人机贴近铁塔和输电线路进行细节巡视的测距和避障飞行提供可靠的依据;同时绝缘子为故障多发元件严重威胁电网的安全,需充分利计算机技术对其进行故障诊断。通过搭建卷积神经网络,在由5个卷积池化模块和2个全连接模块组成的经典架构的基础上,对网络进行改进,实现在复杂航拍背景中绝缘子检测。同时在训练的网络模型中抽取绝缘子的特征融入自组织特征映射网络中实现显著性检测,结合超像素分割和轮廓检测等图像处理方法对绝缘子进行数学建模,提出一种针对绝缘子自爆故障的识别算法,取代人工分析,降低由人为经验判断可能造成的误差。经测试,复杂航拍背景下的绝缘子检测精度达90%以上,自爆识别准确率达到85%以上,均满足工程需求,有效提升巡检的效率和智能化水平。
[Abstract]:Unmanned aerial vehicles (UAVs) can obtain a large number of detailed images through high-definition cameras and graphics transmission devices. Insulators are very important and large components in transmission line. Detecting insulators quickly and accurately in image and video can provide reliable basis for ranging and obstacle avoidance flight of UAV close to iron tower and transmission line. At the same time, the insulator is a fault-prone element that seriously threatens the safety of the power network, which needs to be fully diagnosed by computer technology. Based on the classical architecture of five convolution pool modules and two fully connected modules, the network is improved to detect insulators in complex aerial photography background by building a convolution neural network. At the same time, the feature of insulator extracted from the trained network model is integrated into self-organizing feature mapping network to realize salience detection, and the mathematical model of insulator is established by combining image processing methods such as super-pixel segmentation and contour detection. A fault identification algorithm for insulator self-detonation is proposed, which replaces manual analysis and reduces the error caused by human experience. The test results show that the detection accuracy of insulator in complex aerial photography background is more than 90%, and the accuracy of self-detonation identification is more than 85%, which meets the requirements of engineering and effectively improves the efficiency and intelligence level of patrol inspection.
【作者单位】：电子科技大学自动化工程学院;
【分类号】：TM755;TP183;TP391.41

【相似文献】