[1]翟永杰,王乾铭,杨旭,等.融合外部知识的输电线路多金具解耦检测方法[J].智能系统学报,2022,17(5):980-989.[doi:10.11992/tis.202107026]
ZHAI Yongjie,WANG Qianming,YANG Xu,et al.A multi-fitting decoupling detection method for transmission lines based on external knowledge[J].CAAI Transactions on Intelligent Systems,2022,17(5):980-989.[doi:10.11992/tis.202107026]
点击复制
《智能系统学报》[ISSN 1673-4785/CN 23-1538/TP] 卷:
17
期数:
2022年第5期
页码:
980-989
栏目:
学术论文—智能系统
出版日期:
2022-09-05
- Title:
-
A multi-fitting decoupling detection method for transmission lines based on external knowledge
- 作者:
-
翟永杰1, 王乾铭1, 杨旭1, 赵振兵2, 赵文清1
-
1. 华北电力大学 控制与计算机工程学院, 河北 保定 071003;
2. 华北电力大学 电气与电子工程学院, 河北 保定 071003
- Author(s):
-
ZHAI Yongjie1, WANG Qianming1, YANG Xu1, ZHAO Zhenbing2, ZHAO Wenqing1
-
1. School of Control and Computer Engineering, North China Electric Power University, Baoding 071003, China;
2. School of Electrical and Electronic Engineering, North China Electric Power University, Baoding 071003, China
-
- 关键词:
-
输电线路; 金具; 深度学习; 目标检测; 共现知识; 空间知识; 知识推理模块; 解耦检测
- Keywords:
-
transmission line; fitting; deep learning; object detection; co-occurrence knowledge; spatial knowledge; knowledge reasoning module; decoupling detection
- 分类号:
-
TP18
- DOI:
-
10.11992/tis.202107026
- 文献标志码:
-
2022-05-13
- 摘要:
-
为了有效解决输电线路多金具检测过程中存在的目标密集和目标间相互遮挡的问题,提出了融合外部知识的多目标解耦检测方法(external knowledge decoupling R-CNN, EKD R-CNN)。首先通过深入分析金具数据集的领域规则和图像信息,提取出共现和空间知识;然后使用图神经网络方法建立共现和空间知识推理模型,将外部知识进行实例化表达;最后使用解耦模块将金具检测任务以非耦合的方式进行训练和学习。在实验阶段,对具有14类金具的数据集,进行多种定性和定量实验。对比实验表明,EKD R-CNN的检测效果优于其他先进目标检测模型,与原有基线模型相比,准确率提高6.6%;定性实验表明算法能够解决目标遮挡的问题,实现密集目标的检测;消融实验表明,每种模块对模型的检测效果均有一定的提升。
- Abstract:
-
This paper proposes a multi-object decoupling detection method based on external knowledge (EKD R-CNN) to effectively solve the problems of object density and mutual occlusion in the process of multi-fitting detection of transmission lines. First, the domain rules and image information of the fittings datasets are deeply analyzed to extract the co-occurrence and spatial knowledge; then, graph neural network methods are used to build co-occurrence and spatial knowledge reasoning models, to instantiate and express external knowledge; finally, the decoupling module is employed to train and learn the fittings detection task in an uncoupled way. Multiple qualitative and quantitative experiments are conducted on datasets with 14 types of fittings in the experiment phase. The contrast experiment shows that the detection effect of EKD R-CNN is better than that of other advanced object detection models and that compared with the original baseline model, the detection accuracy of the algorithm is increased by 6.6%; the qualitative experiments suggest that the proposed algorithm can solve the problem of object occlusion, and the ablation experiment indicates that each module improves the detection effect of the model to a certain extent.
备注/Memo
收稿日期:2021-07-15。
基金项目:国家自然科学基金项目(61773160, 61871182);北京市自然科学基金项目(4192055);河北省自然科学基金项目(F2021502013, F2020502009, F2021502008).
作者简介:翟永杰,教授,博士,主要研究方向为电力视觉。主持国家自然科学基金面上项目1项、河北省自然科学基金项目1项,主持横向科研项目多项,参与国家重点研发计划项目1项,获山东省科技进步一等奖1项。授权发明专利10项,编著1部,参编教材1部、著作3部,发表学术论文30余篇;王乾铭,博士研究生,主要研究方向为电力视觉与知识推理;赵振兵,教授,博士,主要研究方向为电力视觉。主持国家自然科学基金等纵向课题10项,获省科技进步一等奖1项(第3完成人)。以第1完成人获得国家专利授权16项,以第1作者出版专著2部,发表学术论文50余篇。
通讯作者:赵振兵. E-mail:zhaozhenbing@ncepu.edu.cn
更新日期/Last Update:
1900-01-01