Fly paddle PaddlePaddle together China Southern Power Grid, electrical inspection towards the "no time"

Guangdong summer, unbearably hot.

When most people enjoy the cool air-conditioned room to bring the power of a group of people braved the heat had to run around to work outdoors, they are the Guangdong Power Grid Corporation of China Southern Power Grid electricity inspection personnel.

Under the scorching sun, in order to protect their own personal safety, electrical inspection personnel need to wear long-sleeved pants cotton overalls, wearing a helmet to conduct an inspection and inspection of outdoor equipment. "Walk around down the body is sweat, clothes can be out of water," Dr. Yang Ying instrument Guangdong Southern Power Grid Energy Technologies Division robot so described the hardships of summer outdoor inspection work.

Guangdong Electric Power Research Institute Energy Technology companies use Baidu fly paddle (PaddlePaddle) deep learning platform for intelligent substation inspection robot independently developed to provide visual energized, accurate detection and analysis of substation equipment, so that the original single 6 hours site visits by robots artificial substitute, greatly reducing operation and maintenance costs and improve the intelligence level of patrol work.

Here Insert Picture Description
To send electricity through from production to consumption, transmission, transformation and distribution, with five link, any link problems, will affect the normal supply of electricity. Wherein the power transmission is an important part of the power transmission through the grid. Electrical inspection of the core content of the work, is to carry out transmission and transformation equipment operation and maintenance, to ensure their normal work to ensure a stable supply of energy and safe operation of the power system.

All along, the electrical inspection is done by hand. Where substation, where there are transmission lines, electrical inspection personnel will go. Whether it is the endless plains, or jungle-clad mountains.

It substation inspection, the inspection of equipment is an important part of it. The road, "We are not in the inspection is to patrol every day: now become a senior engineer Dr. Yang Ying instrument is also to be on the scene patrol dimensional electrical equipment production line too, he recalled where the team's old squad leader joking remark . "
Here Insert Picture DescriptionElectrical equipment inspection work is boring and tedious, a typical substation patrol point patrol work involves up to more than 1,000, which usually takes two staff members spend 6-7 hours to complete, more than labor-intensive, time-consuming and more . With the rapid economic development, increasing the size of the grid, the number of power transmission equipment more and more. Under the current staff size, how to provide intelligent inspection equipment, the less human input, higher operation and maintenance efficiency, it is placed in front of all the people the power of a pressing problem.

Here Insert Picture Description

多年来,广东电网也一直在进行各种方法的尝试。目前,机器人已经被引入电力巡检领域,并可以取代人工完成大部分自动化巡视工作。但是,针对表计识别的传统图像识别方法由于受到复杂背景、光线条件等因素的影响,检测与识读的整体准确率仍然较低。

广东电网通过与百度公司的合作,利用AI算法实现装备赋能升级。

基于飞桨的AI算法帮助变电站智能巡检机器人提高了设备图像识别的准确率,提升了装备的智能化水平,进一步解放了劳动力。

现在,利用智能巡检机器人开展室外巡视,只需预先设定机器人的巡检点,规划好巡检路线,机器人就会自动进行所有相关表计的检测与读取。工作人员无需花费6个小时在现场巡视,而只需要在远方的主控室一键下达巡检任务。

“所以现在我们变电站可以利用智能巡检机器人替代人工开展现场巡视了”,杨英仪说道,智能巡检机器人的好处在于,受环境、气候及作业时长等因素的影响较小,可以降低人工巡检的劳动强度,降低运维成本,提高巡检作业和管理的自动化和智能化水平。从更长远的意义上讲,智能巡检机器人的推广应用与赋能提升,可有效推动变电站巡检无人化的进一步发展。

Here Insert Picture Description
Here Insert Picture Description

新技术的应用并不是一帆风顺的,中间也走了不少弯路。

2016年,广东电网围绕智能巡检机器人研发成立了重点攻关团队,团队中有做高压绝缘的,做传感检测的,但是没有一个人是机器人专业科班出身的。用他们的话说,他们对机器人技术“一抹黑”,完全不了解。经过2年多的实践和摸索,光机器人本体和后台,更迭了至少六个版本。

以变电设备巡视中的指针类表计检测和识读为例,在最开始的时候,他们使用传统的图像处理技术,用人工设计特征方法进行特征提取。这样的方法只能获取图像中目标对象的浅层特征,这使得识别过程极易受到光照等环境因素的影响,在各类干扰因素影响下,整体识别准确率并不高。

Here Insert Picture Description2017年,广东电网与百度建立了战略合作关系,赋能前端巡检设备也是合作的重要内容之一。智能巡检机器人攻关团队引入百度飞桨(PaddlePaddle)平台,利用其所提供的YOLOv3、U-Net,使机器人面向表计的深层次特征提取能力大大提高,突破了环境因素的制约,方法的准确率和鲁棒性显著提升,在表计目标检测、示数读取等方面的效果尤为显著。现在,对于用原有传统方法处理起来极为困难的表计目标,用深度学习方法已经可以获得处理效果的有效提升。

YOLOv3是一个速度和精度均衡的目标检测网络,飞桨物体检测统一框架PaddleDetection通过增加mixup、label_smooth等处理,对YOLOv3进行了优化实现。YOLOv3也是一个单阶段的目标检测器。传统目标检测方法通过两阶段检测,第一阶段生成预选框,第二阶段对预选框进行分类得到类别,而YOLO将目标检测看作是对预测框位置的一个单阶段回归问题。因此,推理速度能够达到具有同样精度的两阶段目标检测方法的几乎2倍。此外,YOLOv3在最初版YOLO的基础上引入了多尺度预测,因而对小物体的检测精度大幅提高。

U-Net是飞桨语义分割库PaddleSeg中支持的四个主流分割网络之一,整个网络是标准的encoder-decoder网络,具有参数少、计算快、应用性强等特点,对场景有较高的适应度。U-Net使用跳跃连接,以拼接的方式将解码器和编码器中相同分辨率的feature map进行特征融合,帮助解码器更好地恢复目标的细节。

在变电站表计示数识读这一案例中,机器人对表计的读取需要经过表计整体目标检测及二次对准、表盘目标检测及示数读取两个阶段。原来的机器人主要是采用基于传统人工设计特征的图像处理方法实现表计的目标匹配和轮廓检测。由于这些方法只能实现浅层特征的提取,在应用的过程中容易受到图像背景、环境光照、拍摄角度等因素的影响,分类错误率较高。此外,样本量的增加对此类方法的作用也不大,大量深层特征无法被挖掘并用于提升算法的性能。

Here Insert Picture DescriptionHere Insert Picture Description
Here Insert Picture Description
Here Insert Picture Description
针对这些问题,对机器人的赋能升级采用了基于深度学习的方法,对涉及目标检测与图像分割的两个关键阶段进行了改善。出于降低图像处理的传输时延和资源需求、提高前端处理效率和智能化程度的目的,上述两个阶段均是在机器人本体上完成。因此,改进方案选择了更适用于前端推理实现的YOLOv3和U-Net模型。

在面向表计目标检测的具体实现中,攻关团队所采用的训练集包含了946张图像样本,标注了2838个表计目标,而测试集则包含了236张图像样本。在基于飞桨YOLOv3的检测实现中,对表盘检测的最高mAP达到了0.9857,有效地提升了目标检测的查全率和查准率。

Here Insert Picture DescriptionHere Insert Picture Description
在面向表计示数识读的具体实现中,利用翻转、旋转、随机裁剪等方式进行图像预处理,为U-Net模型提供更多不同的训练样本;在表盘分割中,则利用U-Net实现表计指针、表盘刻度线、表盘示数等关键要素的有效提取,继而利用圆心定位、数字分类、角度估算等后处理方式实现表计示数的有效估计。在101张测试样本中,指针示数识别结果的偏差与召回率的分布如下表所示。

Here Insert Picture Description

目前,广东电网和百度飞桨的合作主要集中在变电设备巡检领域,未来将在输电线路巡检、现场风险管控等方向上开展合作,逐步推动电力巡检行业向智能化方向发展。

作为中国全面开源开放、功能完备的产业级深度学习平台,百度飞桨已经成为全面推动国内产业智能化升级的重要基石。正如飞桨获得第六届世界互联网颁发的“世界互联网领先科技成果”这份荣誉所彰显的,飞桨技术领先、功能完备、生态丰富等特点向世界展示着中国科技的力量。

Here Insert Picture Description

  • 基于PaddleDetection的YOLOv3模型训练过程可参考:
    https://github.com/PaddlePaddle/PaddleDetection/blob/master/docs/YOLOv3_ENHANCEMENT.md

  • Based on U-Net model training process PaddleSeg may refer to:
    https://github.com/PaddlePaddle/PaddleSeg/blob/release/v0.2.0/turtorial/finetune_unet.md

  • Dial split tutorials Reference:
    https://github.com/PaddlePaddle/PaddleSeg/blob/release/v0.3.0/contrib/README.md#%E5%B7%A5%E4%B8%9A%E7%94%A8 % E8% A1% A8% E5 % 88% 86% E5% 89% B2

Want to connect with more depth learning developer, join the fly paddle official QQ group: 796 771 754.

If you want to learn more about the content flying paddle PaddlePaddle more, see the following document.

  • Official website address: https: //www.paddlepaddle.org.cn/
Published 116 original articles · won praise 1 · views 4585

Guess you like

Origin blog.csdn.net/PaddleLover/article/details/103731575