[1]冯小龙,王 涛,于 欢,等.基于SOM 神经网络的电力设备运行状态评估研究[J].机械与电子,2026,44(07):105-111.
　FENG Xiaolong,WANG Tao,YU Huan,et al.Operational State Assessment of Power Equipment Based on SOM Neural Network[J].Machinery & Electronics,2026,44(07):105-111.

点击复制

基于SOM 神经网络的电力设备运行状态评估研究()

分享到：

《机械与电子》[ISSN:1001-2257/CN:52-1052/TH]

卷:

44

期数:

2026年07期

页码:

105-111

栏目:

电力控制

出版日期:

2026-07-25

文章信息/Info

Title:

Operational State Assessment of Power Equipment Based on SOM Neural Network

文章编号:

1001-2257(2026)07-0105-07

作者:

冯小龙¹; 王 涛¹; 于 欢²; 雍彩晶¹; 高召涛¹

1.国能(甘肃)新能源有限公司,甘肃 白银 730900;

2.国家能源集团甘肃电力有限公司,甘肃 兰州 730030

Author(s):

FENG Xiaolong¹; WANG Tao¹; YU Huan²; YONG Caijing¹; GAO Zhaotao¹

(1.Guoneng (Gansu) New Energy Co.,Ltd.,Baiyin 730900,China;

2.Gansu Electric Power Co.,Ltd.,National Energy Group,Lanzhou 730030,China)

关键词:

电力设备状态评估; 自组织映射; 层次化拓扑学习; 异常预警

Keywords:

power equipment state assessment; self-organizing map; hierarchical topological learning; anomaly early warning

分类号:

TP183;TM732

文献标志码:

A

摘要:

针对电力设备运行状态评估中存在的标签稀缺、多源特征强耦合以及异常渐进演化难以刻画等问题,提出一种基于层次化自组织映射(SOM)的电力设备运行状态评估方法。首先构建3层层次化SOM 结构,通过底层细粒度运行模式学习、中层子系统稳健聚合以及顶层设备级状态重构,实现高维运行特征向低维可解释拓扑空间的逐级映射;其次引入融合激活频次与量化误差的自适应层间权重机制,在保持拓扑关系的同时提升状态表征的统计稳健性与抗噪能力;在健康评估层面,基于设备级结构化表征偏离度构建健康指数模型,并通过分布标定实现跨设备尺度一致的风险量化与提前预警。仿真与对比实验结果表明,所提方法在无监督条件下实现了95.6%的状态识别准确率,并显著提升异常提前预警时间与结果稳定性。

Abstract:

To address the challenges in operational state assessment of power equipment,including label scarcity,strong coupling among multi-source features,and the difficulty in characterizing gradual anomaly evolution,this study proposes a method based on a hierarchical Self-Organizing Map (SOM).A three-layer hierarchical SOM architecture is first constructed,which progressive projects high-dimensional operational features into a low-dimensional interpretable topological space through fine-grained pattern learning at the bottom layer,robust subsystem aggregation at the middle layer,and device-level state reconstruction at the top layer.An adaptive inter-layer weighting mechanism that integrates activation frequency and quantization error is then introduced to enhance the statistical robustness and noise resilience of the state representation while preserving topological relationships.For health evaluation,an exponential health index model is derived from the deviation of device level structured representations,and distribution calibration is employed to enable scale-consistent risk quantification and early warning across different devices.Simulation and comparative experimental results demonstrate that the proposed method achieves a state recognition accuracy of 95.6% under unsupervised conditions,and significantly improves the anomaly early warning lead time and result stability.

备注/Memo

备注/Memo:

收稿日期:2026-03-05

基金项目:甘肃省创新科技项目(E586700018)

作者简介:冯小龙 (1988-),男,甘肃漳县人,助理工程师,研究方向为电力设备在线监测。

常用功能