视听觉协同的电网目标检测网络  

Visual-and-auditory-fusion Power Grid Target Detection Network

作　　者：侯春羽 侯永宏 朱新山 李亚霖 田杨阳 李斌[1] HOU Chunyu;HOU Yonghong;ZHU Xinshan;LI Yalin;TIAN Yangyang;LI Bin(Key Laboratory of Smart Grid of Ministry of Education(Tianjin University),Tianjin 300072,China;State Grid Henan Electric Power Research Institute,Zhengzhou 450052,China)

机构地区：[1]智能电网教育部重点实验室(天津大学),天津300072 [2]国网河南省电力公司电力科学研究院,郑州450052

出　　处：《高电压技术》2024年第9期4048-4057,共10页High Voltage Engineering

基　　金：国家电网有限公司科技项目(面向智能电网运维场景的视听觉主动感知与协同认知技术研究及应用)(5600-2020463 47A-0-0-00)。

摘　　要：随着电网规模的不断扩大和日益复杂,发展智能运维技术是提升运维效率的必由之路。为实现变电站目标的主动感知,基于小样本学习框架提出一种视听觉协同的目标检测网络(visual and auditory fusion detection network, VAFDNet),利用小样本量级数据融合视听觉信息,实现低数据集成本的模型扩展。VAFDNet由作为基础网络的Faster R-CNN和声云信息引导模块构成。声云信息引导模块包括声学特征引导分支和全局特征引导分支,并通过引入声云信息引导机制实现视听觉特征的融合与增强。VAFDNet的训练利用2阶段训练微调方法,提升了网络的泛化性能,可以有效应对变电站目标视听觉协同样本稀缺的问题。在含有3类目标的视听觉协同样本集上进行测试,VAFDNet整体识别精度达到52.623%,各类视听觉协同目标的检测精度均有大幅提升,对数据量极小的主变压器、套管的识别效果也得到明显改善。With the continuously extending and increasing complexity of power grid structure,developing intelligent operation and maintenance technology is the only way to improve the operation and maintenance efficiency.To achieve the active perception of the substation objects,a visual and auditory fusion detection network(VAFDNet)is presented based on a few shot object detection framework,and visual and auditory information is fused with minor level data to im-plement the model extension with low dataset cost.The VAFDNet is constituted by Faster R-CNN as the basic network and the acoustic cloud information guidance module,which includes acoustic feature guidance branch and global feature guidance branch,and the fusion and enhancement of visual and auditory features are realized by introducing acoustic cloud information guidance mechanism.The VAFDNet uses a two-stage fine-tuning training method,improving the gen-eralization performance of the network,which can effectively solve the problem of scarce visual and auditory fusion sample.The VAFDNet network is tested on the dataset of visual and auditory fusion samples with 3 types of targets,and the overall recognition accuracy of the proposed model reaches 52.623%.The detection accuracy of all kinds of visual and auditory fusion targets has been greatly improved,and the identification effect of the transformer and casing with a very small data amount has also been significantly improved.

关 键 词：智能运维 视听觉协同 模型扩展 目标检测 声云信息引导 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TM73[自动化与计算机技术—计算机科学与技术]