面向电力多模态融合的语义差异性和感知能力差异性分析方法  

Semantic Difference and Performance Difference Analysis Method for Power Multimodal Data fusion

作　　者：王红霞 王波[1,2] 董旭柱 姚良忠[1,2] 张嘉鑫 马恒瑞 WANG Hongxia;WANG Bo;DONG Xuzhu;YAO Liangzhong;ZHANG Jiaxin;MA Hengrui(Hubei Key Laboratory of Power Equipment&System Security for Integrated Energy,Wuhan 430072,China;School of Electrical Engineering and Automation,Wuhan University,Wuhan 430074,China)

机构地区：[1]综合能源电力装备及系统安全湖北省重点实验室,武汉430072 [2]武汉大学电气与自动化学院,武汉430074

出　　处：《高电压技术》2024年第9期4037-4047,I0021,共12页High Voltage Engineering

基　　金：国家重点研发计划(2021YFB2401302)。

摘　　要：多模态融合是实现电力系统数字化的重要技术手段,但多模态特征间的差异性限制了融合感知效果。因此,首先对电力多模态数据融合中的语义差异性和感知能力差异性现象进行了深入分析,对差异性产生的特征同化和权重决策问题进行了剖析。然后,针对语义差异性问题,使用角度差对多模态语义差异进行表征,并基于此寻找联合表征空间,实现电力多模态特征同化;其次,针对感知能力差异性问题,使用交叉损失熵对电力多模态感知能力进行表征,并基于此构建权重决策模块,实现多模态特征融合权重计算。最后,以前期所提融合框架为基础,提出了针对多模态差异性问题的高容错性特征融合模型。仿真以输变电线路应急抢修场景为例,基于所提分阶段训练策略进行模型训练,并从融合感知、特征同化以及权重决策机制3个方面验证了所提方法的有效性。Multimodal data fusion is an important technique to realize digitalization of power systems.However,the se-mantic difference caused by data form,physical meanings and so on,and the perception ability difference caused by data incompleteness,perception mechanism and so on restrict its development and application.Taking electrical structured time series and unstructured images as fusion objects,we put forward a feature fusion method with high fault tolerance based on feature assimilation and weight decision-making mechanism.Firstly,different convolution neural networks are used to extract features for all modes of data.Secondly,for semantic difference,angle difference is used to represent mul-ti-modal semantic difference,based on which the joint representation space is searched to achieve the feature assimilation.Thirdly,for perception ability difference,cross loss entropy is used to characterize the target perception ability of different modes,based on which the fusion weight determination mechanism is constructed to calculate fusion weight.Finally,the multi-modal features are concatenated to perceive the power target.The emergency repair scenario of power transmission and distribution lines is taken as an example,and the model is trained based on the phased training strategy.In addition,the effectiveness of the proposed method is verified from three aspects including fusion based perception,feature assimi-lation,and weight decision-making mechanisms.

关 键 词：多模态特征融合 语义差异性 感知能力差异性 特征同化 权重决策 

分 类 号：TM73[电气工程—电力系统及自动化] TP399[自动化与计算机技术—计算机应用技术]