摘要
气体绝缘组合电器(GIS)母线触头的对接深度不足、对中度偏差和弹簧松弛等接触缺陷会显著增加接触电阻,引发过热甚至绝缘击穿事故。针对现有检测方法在复杂工况下灵敏度不足的问题,文中提出一种基于电流—磁场双模态数据融合与交叉注意力机制的缺陷识别算法。通过GIS原型试验平台同步采集4种典型工况下的瞬态电流信号与16通道空间磁场分布数据,构建多物理场数据集;采用Mel滤波器提取电流时频特征,通过对称点模式(SDP)表征磁场空间特性;进而设计双分支ResNet-18网络分别处理两类特征,通过交叉注意力机制动态分配模态权重,实现1024维融合特征的深度提取;最终经多层感知机分类器输出缺陷类型。实验表明:所提模型在10折交叉验证中平均分类准确率达92.76%,较传统拼接融合和双线性池化方法显著提升。t-SNE可视化证实融合特征具有更优的类间分离度,单次推理耗时低于50 ms,满足工程实时性需求,为GIS电接触状态的在线监测与早期预警提供了高精度解决方案。
Such contact defects as insufficient butt depth,alignment deviation and spring relaxion of bus contact of gas insulated switchgear(GIS)will significantly increase contact resistance,leading to overheating and even insulation breakdown accidents.Aiming at the problem of insufficient sensitivity of existing detection methods under complex working conditions,a defect identification algorithm based on current-magnetic field bimodal data fusion and cross-attention mechanism is proposed in this paper.Through GIS prototype test platform,the transient current signal and 16-channel spatial magnetic field distribution data under four typical working conditions are synchronously collected,and a multi-physics dataset is constructed.Mel filter is used to extract the time-frequency characteristics of current,and the spatial characteristics of magnetic field are characterized by symmetric point mode(SDP).Then,a dual-branch ResNet-18 network is designed to process two types of features respectively,and modal weights are dynamically allocated through cross-attention mechanism to achieve deep extraction of 1024-dimensional fusion features.Finally,the defect type is output by the multi-layer perceptron classifier.Experiments show that the average classification accuracy of the proposed model in 10-fold cross-validation is 92.76%,which has significant improvement compared with the traditional stitching and fusion and bilinear pooling methods.The t-SNE visualization proves that the fusion features have better inter-class separation,and the single reasoning time is less than 50 ms,which meets the real-time requirements of engineering and provides a high-precision solution for online monitoring and early warning of electrical contact status of GIS.
作者
谢志杨
彭涛
敬磊
XIE Zhiyang;PENG Tao;JING Lei(Foshan Power Supply Bureau of Guangdong Power Grid Co.,Ltd.,Guangdong Foshan 528000,China)
出处
《高压电器》
北大核心
2025年第10期117-127,共11页
High Voltage Apparatus
基金
南方电网公司科技项目(基于瞬态电磁耦合与分布磁测量的GIS母线电接触状态检测方法研究)。
关键词
GIS母线
电接触缺陷
电流—磁场融合
交叉注意力
深度学习
GIS bus
electrical contact faults
current-magnetic field fusion
cross-attention mechanism
deep learning
作者简介
通信作者:谢志杨(1975-),男,硕士,高工,主要研究方向为高电压技术(E⁃mail:youngwy_001@163.com);彭涛(1973-),男,本科,高工,主要研究方向为变电一次设备技术(E⁃mail:pt_ao@msn.com);敬磊(1988-),男,本科,工程师,主要研究方向为高压开关设备技术(E⁃mail:fsgdjl@163.com)。
