地下管廊巡检机器人的环境感知与自主导航技术研究

CASIA OpenIR > 毕业生 > 硕士学位论文

	地下管廊巡检机器人的环境感知与自主导航技术研究
	闫帅
	2017
学位类型	工程硕士
中文摘要	现代社会城市用地紧张、交通压力大、市容建设要求高，地下电缆因而成为城市或人口密集区电网系统的主要输电方式，其可靠稳定运行是电网安全和高质量供电的基础。目前地下电缆的巡检仍以人工为主，地下环境条件恶劣，巡检劳动强度大，危险性高，尚缺乏行之有效的自动化巡检手段。随着机器人技术的迅猛发展及其在不同应用领域的不断深入，将移动机器人应用于地下输电线路的实时监测中是一个必然的发展趋势。机器人进入电缆管廊中进行检测作业，可大幅降低工作人员的劳动强度，提高作业的工作效率，避免人工作业的危险，减少电缆故障和电缆沟事故的发生概率。因此，输电管廊巡检机器人的研究对于提高输电质量、保障输电安全具有重要的应用价值和实际意义。本文以提高无轨式管廊移动巡检机器人的整体控制性能为目标，开展了管廊机器人巡检课题中的关键技术的研究工作，主要是环境感知检测算法的研究和机器人导航控制算法的研究。论文的主要工作如下： 1、提出了一种基于深度学习的关键点检测算法，并将其应用于道路消失点的检测问题中。基于AlexNet网络，结合消失点检测问题需求修改了网络结构，替换了激活函数，增加了一层全连层，修改了loss函数，将其应用于消失点的检测。实验结果表明，本文提出的算法适应性很强，对于传统方法很难处理的模糊图片依然有很好的效果。 2、采用FCN网络对管廊内路面的分割问题开展了研究。针对地下管廊的特点简化了FCN的网络结构，并加入了代表消失点位置的关键层。算法将消失点的信息综合到路面分割中，用消失点的位置为路面分割做出指导，实现了路面的像素级分类。实验表明，该算法对于地下管廊环境相较于普通深度学习分割算法表现的更好。 3、实现了巡检机器人仿真环境和真实环境的转换算法。基于变分自编码器，对抗生成算法，和循环对抗生成算法，搭建了无监督图像转换的网络，实现了能够将机器人仿真环境和机器人真实环境相互转换的算法。这将使得基于学习算法的视觉导航能够更快适应真实场景。 4、采用强化学习算法实现了移动机器人在管廊环境中的路径导航。通过建立一个仿真环境，把激光雷达的输入作为状态，把智能体的输出作为机器人的转向控制，通过智能体的深度Q-learning算法，实现了仿真环境下机器人的自主导航。机器人通过在环境中不断的学习，找到不同状态下的最优动作。该算法具有能自动获取环境知识、自适应能力强等优点。
英文摘要	In modern society, the urban land is tense, the traffic pressure is great, and the demand for city appearance is high. The underground cable thus becomes the main transmission method of the power grid system in the urban or densely populated areas. Its reliable and stable operation is the basis for power grid security and high-quality power supply. At present, the inspection of underground cables is still dominated by manpower, underground environmental conditions are poor, labor intensity of inspections is large, and the danger is high. There is still no effective automated inspection method. With the rapid development of robotics technology and its in-depth application in different fields, it is an inevitable trend to apply mobile robots to the real-time monitoring of underground transmission lines. The robot enters the cable duct for inspection operations, which can greatly reduce the labor intensity of the workers, improve the work efficiency, avoid the danger of manual operations, and reduce the probability of cable faults and cable trench accidents. Therefore, the research on the inspection robot of power transmission pipelines has important application value and practical significance for improving the transmission quality and ensuring the transmission safety. This paper aims at improving the overall control performance of the trackless pipeline inspection mobile robot, and carries out the research work on the key technologies in inspection task of the pipeline robot. The mainly researches are environment perception algorithm and the robot navigation control algorithm. The main work of the dissertation is as follows: 1、A key point detection algorithm based on deep learning is proposed and applied to the problem of road vanishing point detection. Based on the AlexNet network, combined with the vanishing point detection problem, the network structure was modified, the activation function was replaced, an additional layer was added, the loss function was modified, and it was applied to the detection of vanishing points. The experimental results show that the method proposed in this paper is highly adaptable, and it still has a good effect on fuzzy pictures that are difficult to handle with traditional methods. 2、The FCN network was used to study the segmentation of the pavement in the pipeline. According to the characteristics of the underground gallery, the FCN's network structure was simplified, and a key layer representing the vanishing point location was added . The algorithm integrates the vanishing point information into the road segmentation, uses the vanishing point position to guide the road segmentation, and realizes the pixel level classification of the road surface. Experiments show that this method performs better for the underground gallery environment than the ordinary deep learning segmentation algorithm. 3、The conversion algorithm for inspection robot simulation environment and real environment was proposed and implemented. Based on variational auto-encoder, generative adversarial algorithm, and loop generative adversarial algorithm, a network of unsupervised image conversion was established, and an algorithm capable of converting the robot simulation environment and the robot's real environment was realized. This will enable visual navigation based on learning algorithms to adapt more quickly to real scenes. 4、The reinforcement learning method is used to achieve the path navigation of mobile robots in the corridor environment. By establishing a simulation environment, the input of the laser radar is taken as the status, and the output of the agent is used as the steering control of the robot. The algorithm realizes autonomous navigation of the robot under the simulation environment through deep Q-learning algorithm. Through continuous learning in the environment, robots find the best actions in different states. This will allow visual navigation based on learning algorithms to adapt to real-world situations faster.
关键词	管廊机器人巡检消失点分割导航
文献类型	学位论文
条目标识符	http://ir.ia.ac.cn/handle/173211/21480
专题	毕业生_硕士学位论文
作者单位	中国科学院自动化研究所
第一作者单位	中国科学院自动化研究所
推荐引用方式 GB/T 7714	闫帅. 地下管廊巡检机器人的环境感知与自主导航技术研究[D]. 北京. 中国科学院研究生院,2017.

条目包含的文件
文件名称/大小	文献类型	版本类型	开放类型	使用许可
闫帅-毕业论文.pdf（4375KB）	学位论文		限制开放	CC BY-NC-SA