[1]马 泉,王 皓,凌 扬,等.弱监督时序学习融合物理约束的变压器绝缘弱退化识别[J].机械与电子,2026,44(04):119-126.
　MA Quan,WANG Hao,LING Yang,et al.A Physics-guided Weakly Supervised Temporal Learning for Identifying Subtle Insulation Degradation in Power Transformers[J].Machinery & Electronics,2026,44(04):119-126.
点击复制

弱监督时序学习融合物理约束的变压器绝缘弱退化识别()
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:
119-126
栏目:
电力控制
出版日期:
2026-04-27

文章信息/Info

Title:
A Physics-guided Weakly Supervised Temporal Learning for Identifying Subtle Insulation Degradation in Power Transformers
文章编号:
1001-2257 ( 2026 ) 04-0119-08
作者:
马 泉; 王 皓; 凌 扬; 董晓岽; 周 鑫; 孙伟楠
江苏省送变电有限公司,江苏 南京 210028
Author(s):
MA Quan ;  WANG Hao ;  LING Yang ;  DONG Xiaodong ;  ZHOU Xin ;  SUN Weinan
( Jiangsu Power Transmission and Transformation Co. , Ltd. , Nanjing 210028 , China )
关键词:
变压器实验; 介质损耗; 对比学习; 度量学习; 物理约束; 异常检测
Keywords:
power transformer testing ;  dielectric dissipation factor ;  contrastive learning ;  metric learning ;  physics constraints ;  anomaly detection
分类号:
TM41 ;TP391.4
文献标志码:
A
摘要:
提出一种融合物理约束的弱监督时序特征学习模型( PG-TSML ),通过低频、小样本和短时序特性的例行实验数据,对变压器绝缘老化引起的退化进行识别。该方法通过滑动窗口构造时序样本以强化年度趋势特征,采用轻量时序卷积网络( TCN )作为编码器提取设备健康表征;引入对比学习,通过构造正负样本,基于信息噪声对比估计损失,增强健康特征紧致性;结合度量学习在每个训练批次计算相邻年度表征差异并加到损失上,学习趋势规律;嵌入温度一致性与绝缘老化单调性的物理约束,使模型学习绝缘退化的物理规律;最后通过一类支持向量机进行无监督异常检测,预警评分超阈值时发出早期预警。以 20 台主变5 a 的介损与电容量数据为基础,与 4 类基线方法开展对比实验。结果表明, PG-TSML 的受试者工作特征曲线下面积与 F1 分数分别达到 0.910 和 0.820 ,较传统阈值法提升 0.290 和 0.340 ,提前预警时间达 2.3 a ,误报率仅 0.040 。消融实验表明,对比学习可提升表征紧凑性缓解样本稀缺,度量学习增强对缓慢退化的敏感度并提前预警时间,物理约束保障符合绝缘退化规律降低误报率, TCN 编码器有效捕捉时序依赖和弱退化演化特征。所提方法可在有限样本下有效提取介损演化特征,显著增强变压器绝缘弱退化早期识别能力,具有重要工程应用价值。
Abstract:
This paper proposes a physics guided weakly supervised temporal feature learning model ( PG-TSML ) for the early detection of insulation degradation in power transformers , using routine test data characterized by low frequency , small sample size , and short time series.The method constructs temporal samples through a sliding window to enhance annual trend features , and employs a lightweight temporal convolutional network ( TCN ) as an encoder to extract equipment health representations.Contrastive learning is introduced to improve the compactness of health features by constructing positive and negative pairs under the Info Noise Contrastive Estimation ( InfoNCE ) loss.Metric learning is incorporated by calculating the difference between representations of adjacent years within each training batch and adding it to the loss function , facilitating the learning of trend patterns.Physical constraints , including temperature consistency and the monotonicity of insulation aging , are embedded to guide the model in learning the underlying physical principles of insulation degradation.Finally , an unsupervised One Class Support Vector Machine is then applied for anomaly detection , triggering early warnings when the health score exceeds a predefined threshold.Based on five years of dielectric loss and capacitance data from 20 main transformers , comparative experiments are conducted with four types of baseline methods.The results demonstrate that PG-TSML achieves an area under the ROC curve of 0.910 and an F1 score of 0.820 , representing improvements of 0.290 and 0.340 over the conventional threshold based methods , respectively.It provides an average early warning lead time of 2.3 years with a false alarm rate of only 0.040.Ablation studies show that contrastive learning enhances feature compactness under limited data , metric learning improves sensitivity to gradual degradation and extends early warning time , physical constraints ensure adherence to insulation aging laws and reduce false alarms , and the TCN encoder effectively captures temporal dependencies and weak degradation evolution.The proposed approach enables robust extraction of dissipation factor evolution features from limited samples , substantially improving the early identification of weak insulation degradation in transformers , demonstrating substantial value for engineering applications.

参考文献/References:

[ 1 ] 周家玉,侯慧娟,盛戈皞,等 . 状态参量关联规则挖掘及深度学习融合的变压器故障诊断算法[ J ] . 高压电器,2023 , 59 ( 3 ): 108-115.

[ 2 ] 国家市场监督管理总局 . 油浸式电力变压器技术参数和要求: GB / T6451 — 2023 [ S ] . 北 京:中 国标 准出 版社,2023.
[ 3 ] 国家能源局 . 电力设备预防性实验规程: DL / T 596 —2021 [ S ] . 北京:中国电力出版社, 2021.
[ 4 ] 黄柯予 . 变压器绕组介损实验数据分析系统的设计与实现[ D ] . 成都:电子科技大学,2021.
[ 5 ] 张鸿儒 . 基于数据信息挖掘的电力变压器故障诊断及健康评估[ D ] . 济南:山东大学,2023.
[ 6 ] 谢长宁,史宗尚 . 基于扩展隔离森林算法的小型水力发电系统故障检测研究[ J ] . 机械与电子, 2025 , 43 ( 9 ): 45-50.
[ 7 ] KUMAR A , KUMAR A , RAJA R , et al.Revolutionising anomaly detection : a hybrid framework for anomaly detection integrating isolation forest , autoencoder , and Conv.LSTM [ J ] .Knowledge and information systems , 2025 , 67 : 11903-11953.
[ 8 ] 丁世飞,孙玉婷,梁志贞,等 . 弱监督场景下的支持向量机算法综述[ J ] . 计算机学报, 2024 , 47 ( 5 ): 987-1009.
[ 9 ] NALEPA J , KAWULOK M.Selecting training sets for support vector machines : a review [ J ] .Artificial Intelligence Review , 2019 , 52 : 857-900.
[ 10 ] MALHOTRA P , RAMAKRISHNAN A , ANAND G , et al.LSTM-based encoder-decoder for multi-sensor anomaly detection [ C ] ∥Proceedings of the 2016 IEEE International Conference on Data Science and Advanced Analytics.New York : IEEE , 2016 : 38-43.
[ 11 ] ZHANG H , SUN H L , KANG L , et al.Prediction of health level of multiform lithium sulfur batteries based on incremental capacity analysis and an improved LSTM [ J ] .Protection and control of modern power systems , 2024 , 9 ( 2 ): 21-31.
[ 12 ] 林苑,赵晋斌,孙明琦,等 . 新型电力系统下基于物理信息 LSTM 网络的电力变压器状态评估方法[ J ] . 电力系统保护与控制,2025 , 53 ( 14 ): 133-141.
[ 13 ] ZHANG X Y , ZHANG C , HE R , et al.A pyramidal attention-based transformer model based on improved differential innovation search algorithm and feature extraction for solar radiation prediction considering relevant factors [ J ] .Renewable energy , 2025 , 253 : 123666.
[ 14 ] 张凤,刘雄飞,魏金花 . 基于改进 Transformer 和任务 感知强 化 学 习 的 变 压 器 故 障 预 测 [ J ] . 电 子 器 件,2025 , 48 ( 3 ): 586-592.
[ 15 ] 蒲天骄,乔骥,韩笑,等 . 人工智能技术在电力设备运维检修中的 研究 及 应用 [ J ] . 高 电 压 技 术,2020 , 46( 2 ): 369-383.
[ 16 ] 张志昂,廖光忠 . 改进变分自编码器的工业时序数据异常检测[ J ] . 计算机工程与设计, 2024 , 45 ( 1 ): 17-23.
[ 17 ] 王紫祎,陈世平 . 基于图对比学习的自监督网络流量检测模型[ J ] . 电子科技, 2025 , 38 ( 3 ): 22-31.
[ 18 ] LI D W , TIAN Y J.Survey and experimental study on metric learning methods [ J ] .Neural networks , 2018 , 105 : 447-462.
[ 19 ] KARNIADAKIS G E , KEVREKIDIS I G , LU L , et al.Physics-informed machine learning [ J ] .Nature reviews physics , 2021 , 3 : 422-440.
[ 20 ] 胡海洋,张力,李忠金 . 融合自编码器和 one-class-SVM 的异常事件检测[ J ] . 中国图象图形学报, 2020 ,25 ( 12 ): 2614-2629.
[ 21 ] HERNANDEZ R D , LACHMAN M F.Refining bushing power factor and capacitance analysis through statistics [ J ] .Transformers magazine.2020 , 7 ( 3 ): 20-28.

备注/Memo

备注/Memo:
收稿日期: 2025 12 09
基金项目:江苏省送变电有限公司科技项目( 202405 )
作者简介:马 泉 ( 1981- ),男,江苏南京人,硕士,高级工程师,研究方向为输变电工程调试技术;王 皓 ( 1991- ),男,安徽南陵人,工程师,研究方向为输变电工程实验技术。
更新日期/Last Update: 2026-08-21