[1]陈克洋,谢才科,王嘉昊,等.基于特征注意力与多任务学习的柱上开关健康状态智能评估[J].机械与电子,2026,44(04):106-113.
　CHEN Keyang,XIE Caike,WANG Jiahao,et al.Intelligent Health Status Evaluation of Pole-mounted Switches Based on Feature Attention and Multi-task Learning[J].Machinery & Electronics,2026,44(04):106-113.
点击复制

基于特征注意力与多任务学习的柱上开关健康状态智能评估()
分享到：

《机械与电子》[ISSN:1001-2257/CN:52-1052/TH]

卷:
44
期数:
2026年04期
页码:
106-113
栏目:
电力控制
出版日期:
2026-04-27

文章信息/Info

Title:
Intelligent Health Status Evaluation of Pole-mounted Switches Based on Feature Attention and Multi-task Learning
文章编号:
1001-2257 ( 2026 ) 04-0106-08
作者:
陈克洋; 谢才科; 王嘉昊; 宁 楠; 周依然; 杨政校; 范俊秋
贵州电网有限责任公司,贵州 贵阳 550002
Author(s):
CHEN Keyang ;  XIE Caike ;  WANG Jiahao ;  NING Nan ;  ZHOU Yiran ;  YANG Zhengxiao ;  FAN Junqiu
( Guizhou Power Grid Co. , Ltd. , Guiyang 550002 , China )
关键词:
柱上开关; 状态评估; 特征注意力机制; 多任务学习; 健康指数
Keywords:
pole-mounted switch ;  condition assessment ;  feature attention mechanism ;  multi-task learning ;  health index
分类号:
TM561
文献标志码:
A
摘要:
针对传统柱上开关健康评估方法依赖专家经验、主观性强、微弱故障特征识别能力弱和泛化性能不足,难以适配智能电网大规模运维需求的问题,提出一种融合特征注意力机制与多任务学习的柱上开关健康评估模型。首先,构建在线监测 离线试验 历史台账多源数据体系,提取 17 维特征参数,通过 3σ 准则剔除异常值、Z-Score 标准化消除量纲影响,完成数据预处理;其次,设计卷积特征提取 联合注意力加权 特征融合核心特征层,动态强化绝缘电阻、分合闸线圈电流等关键故障特征的表达;最后,通过加权联合损失函数实现模型协同优化,构建良好、一般、异常、严重四级健康状态评级任务框架。基于某省级电网 237 台柱上开关 5 a 运维数据的实验表明,模型健康指数预测决定系数 R2达 0.991 1 ,状态分类准确率达95.47% ,预测偏差均值仅为 -0.003 3 ,稳定性显著优于传统方法。注意力权重可视化结果证实,模型可自主聚焦绝缘电阻、开关触头温度等核心特征,从数据驱动角度解决了传统模型依赖主观赋权的问题。
Abstract:
To address the limitations of traditional health assessment methods for pole mounted switches , such as strong reliance on expert experience , high subjectivity , weak ability to identify subtle fault features , and insufficient generalization performance , which hinder their adaptability to the large scale operation and maintenance requirements of smart grids , this paper proposes an intelligent health assessment model that integrates a feature attention mechanism with multi-task learning.First , a multi-source data system of “ online monitoring – offline testing – historical records ” is constructed.Seventeen dimensional feature parameters are extracted , and data preprocessing is completed by eliminating outliers using the 3σ criterion and removing dimensional influences through Z-score standardization.Second , a core feature layer of “ convolutional feature extraction–joint attention weighting – feature fusion ” is designed to dynamically enhance the expression of critical fault features such as insulation resistance and opening / closing coil currents.Finally , the model is collaboratively optimized through a weighted joint loss function , and a four level health status rating task framework of “ Good / General / Abnormal / Severe ” is constructed.Experiment results based on five year operation and maintenance data from 237 pole mounted switches within a provincial power grid show that : the model achieves a coefficient of determination ( R2 ) of 0.991 1 for health index prediction , a state classification accuracy of 95.47% , and a mean prediction deviation of merely -0.003 3 , exhibiting significantly superior stability compared to traditional methods.The visualization results of attention weights confirm that the model can autonomously focus on core features such as insulation resistance and switch contact temperature , thereby addressing the issue of subjective weighting inherent in conventional models from a data-driven perspective.

参考文献/References:

[ 1 ] CHEN H , CAI J H , LI Y N , et al.Research on the transmission relationship of grid science and technology innovation indicators under the smart grid [ C ] ∥2023 2nd International Conference on Smart Grids and Energy Systems ( SGES ) .New York : IEEE , 2023 : 19-22.

[ 2 ] ZHANG C , CHEN S Y , JIA K Q , et al.Development and application of 10 kV pole-mounted fully insulated and visualized load switch [ C ] ∥2022 China International Conference on Electricity Distribution ( CICED ) . New York : IEEE , 2023 : 293-297.
[ 3 ] 宋扬,廖志华,杨帆,等 . 绝缘杆作业法带电加装 10 kV 柱上隔离开关绝缘护套技术探讨[ J ] . 电工技术, 2020 ( 23 ): 96-98.
[ 4 ] 肖阳,曾景原,许月惟 . 柱上开关控制器的技术发展及功能应用[ J ] . 大众用电, 2023 , 38 ( 9 ): 35-36.
[ 5 ] WISCHKAEMPER J A , BENNER C L , DON RUSSELL B.A new monitoring architecture for distribution feeder health monitoring , asset management , and real-time situational awareness [ C ] ∥2012 IEEE PES Innovative Smart Grid Technologies ( ISGT ) .New York : IEEE , 2012 : 1-7.
[ 6 ] 黄智鹏,熊洽,陈家超,等 . 基于改进层次分析法的柱上开关成套设备状态评估研究[ J ] . 广东电力,2020 , 33( 3 ): 88-95.
[ 7 ] LIU L , YANG J , HE Z , et al.Research on comprehensive evaluation weights of metering asset operation benefits based on AHP-entropy weight method [ C ] ∥ 2025 4th International Conference on Energy and Electrical Power Systems ( ICEEPS ), 2025 : 283-287.
[ 8 ] 赵晓明,余志慧,张雪松 . 考虑隶属度可变的继电保护变权模糊综合状态评价方法[ J ] . 电力系统保护与控制,2017 , 45 ( 15 ): 22-29.
[ 9 ] CHEN L A , CHEN Z Y , PAN K X , et al.Design of intelligent pole mounted switch controller based on embedded operating system [ C ] ∥2013 26th IEEE Canadian Conference on Electrical and Computer Engineering ( CCECE ) .New York : IEEE , 2013 : 1-4.
[ 10 ] 林昱奂,胡嘉铭,戴伟力,等 . 一种智能柱上开关状态监测与故障诊断系统[ J ] . 环境技术,2020 , 38 ( 6 ):134-138.
[ 11 ] WU X K , YUAN L , FAN J Q , et al.Intelligent fault detection and self-diagnosis module design for pole-mounted switches based on multi-signal fusion [ C ] ∥2025 4th International Conference on Smart Grids and Energy Systems ( SGES ) .New York : IEEE , 2025 : 520-525.
[ 12 ] 赵科,王飞,杨元威,等 . 基于信号特征融合与优化的高压断路器 机械 状 态评 估[ J ] . 高 压 电 器, 2018 ,54( 4 ): 14-19.
[ 13 ] 李聪 . 基于分合闸线圈电流特征的断路器机械故障诊断研究[ D ] . 北京:华北电力大学,2019.
[ 14 ] 赵子强,张孝远,邵锐,等 . 高抗机械健康状态不确定性量化 混 合 预 测 模 型 [ J / OL ] . 振 动、测 试 与 诊 断,2025 : 1-8 [ 2026-01-28 ] .https : ∥link.cnki.net / urlid / 32.1361.V.20250908.0830.002.
[ 15 ] 杨艳红,卢鑫,张雷杰,等 . 考虑智能体前置状态及环境特征自适应机制的强化学习电网调度方法[ J ] . 高电压技术,2024 , 50 ( 8 ): 3497-3509.
[ 16 ] 黄旭锐,于丰源,杨波,等 . 基于 Transformer 网络和多任务学习的园区综合能源系统电 热短期负荷预测方法[ J ] . 南方电网技术, 2023 , 17 ( 1 ): 152-160.
[ 17 ] 罗 浚,施 伟 锋,薛 婷 . 融 合 高 效 局 部 注 意 力 机 制 和KAN-ResNet 的船舶电网短路故障诊断[ J / OL ] . 电力系统及其自动化学报,2025 : 1-11 [ 2026-01-28 ] .https : ∥doi.org / 10.19635 / j.cnki.csu epsa.001711.
[ 18 ] 吕杰,杜振丹,邓佳雯,等 .FFANet :特征融合与注意力机制结合的火烧迹地提取方法[ J ] . 地球信息科学学报,2025 , 27 ( 12 ): 2983-2995.
[ 19 ] 汪 珍珍 . 基 于 机器 学习 的乳 腺 癌 预 诊 断 方 法 研 究[ D ] . 长沙:中南林业科技大学,2024.
[ 20 ] 潘登 . 基于多模态数据的心理健康测评模型研究与系统实现[ D ] . 扬州:扬州大学,2025.
[ 21 ] 张耀辉,王耀坤 . 基于 CSRT 参数化与深度学习的低雷诺数翼型优化方法[ J / OL ] . 北京航空航天大学学报,2025 : 1-18 [ 2026-01-28 ] .https : ∥doi.org / 10.13700 / j.bh.1001 5965.2025.0610.
[ 22 ] 李亦然 . 基于深度学习的微博三孩生育意愿的文本情感分析[ D ] . 济南:山东大学,2023.

相似文献/References:

[1]李鸿奎,程昭龙,周 蕾,等.智能变电站设备管控大数据分析系统研究[J].机械与电子,2020,(10):72.
　LI Hongkui,CHENG Zhaolong,ZHOU Lei,et al.Big Data Analysis System of Intelligent Substation Equipment Management and Control[J].Machinery & Electronics,2020,(04):72.
[2]李 鹏,黄亦翔,夏鹏程,等. 基于一维卷积长短时记忆网络的多信号融合刀具磨损评估 [J].机械与电子,2021,(05):8.
　LI Peng,HUANG Yixiang,XIA Pengcheng,et al.Multi-signal Fusion Assessment of Tool Wear Based on １DCNN-LSTM[J].Machinery & Electronics,2021,(04):8.

备注/Memo

备注/Memo:
收稿日期: 2025-12-22
基金项目:中国南方电网有限责任公司一般科技项目( 061000KC23100007 )
作者简介:陈克洋 ( 1996- ),男,贵州遵义人,工学学士,研究方向为变电运行智能化;范俊秋 ( 1991- ),男,安徽枞阳人,博士,研究方向为新型电力系统调度及关键设备,通信作者, E-mail : fans _ edu _ gzdx@yeah.net
更新日期/Last Update: 2026-08-21