复合式机器人电力巡检与线路表观缺陷检测研究
边疆
Subtype博士
Thesis Advisor谭民 ; 赵晓光
2020-05-27
Degree Grantor中国科学院大学
Place of Conferral中国科学院自动化研究所
Degree Name工学博士
Degree Discipline控制理论与控制工程
Keyword高压架空输电线路巡检 无人机 悬挂巡线小车 巡检环境SLAM 图像质量评估 电力设备目标检测 表观缺陷检测
Abstract

智能机器人高压架空输电线路巡检具有重要的研究价值和广阔的应用前景。本文研制了一种新型复合式架空线路巡检机器人,研究了线路表观缺陷的检测方法。主要研究工作及创新点如下:

第一,针对目前巡线中悬挂式机器人越障困难、装卸繁琐与混合式机器人机构冗余、运动不稳等问题,提出并实现了一种新型高压架空输电线路复合式智能巡检机器人系统和新的巡检方式。该系统由优势互补的无人机和悬挂巡线小车构成,无人机可自主投放与回收小车实现上下地线,小车与飞行机构分离,可稳定精细巡线。户外实验验证了系统与方法的有效性。

第二,针对无人机巡检杆塔及其周边高压电力设备的任务,提出了一种基于点线特征和半稠密重建的视觉SLAM框架。在该框架中,设计了一种针对杆塔结构的快速启发式直线特征提取与匹配方法;推导了线段重建过程与优化点线特征和无人机运动的雅可比矩阵解析表达;给出了GPS数据融合和半稠密重建方法。所提SLAM框架提高了绕塔巡检时无人机定位的精度,环境建图的快速准确性和对杆塔细节结构恢复的能力。仿真和野外实验验证了所提框架的有效性。

第三,为了筛选电力巡检图像以提高样本集有效性,提出了一种基于子区域特征提取的真实退化无参考图像质量评估方法。在该方法中,通过聚焦特征和加权预测,解决了真实退化差异大和子区域无真值的问题;设计了一种巡检退化标注方法和基于评估不确定度的损失函数,解决了主观评估波动的问题。针对通用目标检测方法应用于高压电力设备检测中出现的背景干扰大、速度慢和准确率低等问题,提出了一种面向多类电力设备的高效旋转矩形框检测方法。在制作的数据集与公开数据集上,验证了所提方法的有效性。

第四,针对通用图像表观异常检测,电力设备和架空地线表观缺陷检测的问题,提出了一种基于特征提取与图像重建的半监督表观缺陷检测方法。在该方法中,推导构建了一种条件变分自编码器,使得隐藏空间成为连续和解耦的流形,能够完成自适应聚类和条件生成;设计了一种基于特征判别器的最大化互信息方法,解决了缺陷信息不明确时,正常与异常区分困难的问题;基于结构相似性损失和易收敛生成对抗网络的细粒度重建,设计了一种可度量细小缺陷的异常分数;设计了一种结合隐藏空间正则程度的主动负训练方法。在制作的数据集与公开数据集上,验证了所提方法的有效性。

最后,对本文的工作进行了总结,并指出了需要进一步开展的工作。

Other Abstract

Intelligent robot high-voltage overhead transmission line inspection has important research values and broad application prospects. In this thesis, a new type of cooperative robot is developed for power line inspection and the detection method of surface defects along the power lines is studied. The main contributions of the thesis are as follows:

Firstly, the climbing inspection robots have limited obstacle avoidance capabilities, and difficulty in getting on and off the power line. The hybrid inspection robots have redundant mechanisms and unstable movements. To solve the weaknesses above, a new type of intelligent cooperative inspection robot system for high-voltage overhead transmission lines and a new inspection scheme are proposed and implemented. The system, which has complementary advantages, consists of a UAV (unmanned aerial vehicle) and a climbing inspection trolley. The UAV can autonomously load and unload the trolley on the OGW (overhead ground wires). The trolley can be separated from the flight mechanism and inspect the OGW finely and steadily. Outdoor experiments demonstrate the effectiveness of the system and approaches.

Secondly, for the tasks of UAV inspection for a tower and surrounding high-voltage power equipments, a visual SLAM (simultaneous localization and mapping) framework based on point-line features and semi-dense reconstruction is proposed. In this framework, a fast heuristic line feature extraction and matching method for the transmission tower structure is designed; The reconstruction process of line segments and the analytical expressions of Jacobian matrices for the optimization of point-line characteristics and UAV motion are derived; A method of GPS (global position system) data fusion and a semi-dense reconstruction approach are presented. The SLAM framework improves the accuracy of UAV positioning, the accuracy and the speed of environmental mapping and the ability to recover the detailed structures of tower. Simulation and field experiments verify the effectiveness of the proposed framework.

Thirdly, in order to select power line inspection images to improve the effectiveness of sample sets, a method based on subregion feature extraction for no-reference quality assessment of real degraded images is proposed. In this method, the problems of large differences among real degradations and no ground truths in subregions of an image are solved by the weighted prediction and focusing on the features; An inspection image quality labelling method and a loss function based on evaluation uncertainty are designed to solve the fluctuation problem of subjective evaluations. For the problems of large background interferences, slow speed and low accuracy when the general object detection method is applied to the detection of high-voltage power equipments, an efficient rotating bounding box detection method for multi-type power equipments is proposed. Experiments on the self-built datasets and the public dataset verify the effectiveness of the proposed approaches.

Fourthly, for the general image anomaly detection and the surface defect detection of the high-voltage power equipments and the OGW, a semi-supervised method based on feature extraction and image reconstruction is proposed. In this method, a conditional variational auto-encoder is derived and realized, makes the latent space become continuous and decoupled, enables the adaptive clustering and conditional generation; A method of maximizing mutual information based on feature discriminator is designed to solve the problem of distinguishing normal from anomalies when defect information is not clear; With a fine-grained reconstruction method which is based on the SSIM (structural similarity index) loss and the easily convergent GAN (generative adversarial network), an anomaly score to solve the problem of small surface defect detection is designed; An active negative training method combining the regularization degree in the latent space is designed. Experiments on the self-built datasets and the public datasets verify the effectiveness of the proposed approaches.

Finally, the thesis is summarized and the further works are pointed out.

Pages157
Language中文
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/39030
Collection复杂系统管理与控制国家重点实验室_机器人理论与应用
复杂系统管理与控制国家重点实验室_先进机器人
Corresponding Author边疆
Recommended Citation
GB/T 7714
边疆. 复合式机器人电力巡检与线路表观缺陷检测研究[D]. 中国科学院自动化研究所. 中国科学院大学,2020.
Files in This Item:
File Name/Size DocType Version Access License
Thesis.pdf(26548KB)学位论文 开放获取CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[边疆]'s Articles
Baidu academic
Similar articles in Baidu academic
[边疆]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[边疆]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.