基于胶囊神经网络的电力变压器故障诊断方法研究  

Research on Fault Diagnosis Method of Power Transformer Based on Capsule Networks

作　　者：罗文萱 LUO Wenxuan(International Education Institute,North China Electric Power University,Hebei Baoding 071003,China)

机构地区：[1]华北电力大学国际教育学院,河北保定071003

出　　处：《高压电器》2024年第5期92-98,共7页High Voltage Apparatus

基　　金：国家自然科学基金面上项目(62373148);中央高校基本科研业务费专项资金面上项目(2021MS018)。

摘　　要：电力变压器作为电力系统运行的枢纽设备,其运行状况直接关系到整个系统的运行安全。长期以来,油中溶解气体分析(dissolved gas analysis,DGA)技术由于不易受变压器内部复杂电磁场及外部噪声影响,成为变压器故障诊断的一种有效手段。为了解决现有基于DGA的故障诊断方法的局限性并进一步提升故障诊断的准确率,文中提出了一种基于胶囊神经网络(capsule network,CapsNet)的油浸式电力变压器故障诊断方法。通过结合CapsNet在处理向量数据方面的优势和变压器油中溶解气体的特征,所提方法能够准确地模拟油中溶解气体与故障类型之间复杂的非线性关系,通过动态路由技术和反向传播算法对模型的训练实现关键特征的自动提取。基于电网真实DGA数据的测试结果表明所提方法在故障诊断准确率、宏平均和召回率几何平均3个性能指标上得分数分别为90.48%、82.46%、87.93%,均优于支持向量机(support vector machine,SVM)和深度信念网络(deep belief network,DBN),证明了其在实际变压器故障诊断应用中的优势。The operation status of power transformers,the hub equipment in the operation of electrical power systems,is directly related to the safe operation of the entire system.For a long time,dissolved gas analysis(DGA)has served as an effective means for diagnosing transformer faults mainly due to its immunity to complex electromagnetic fields inside the transformers and external noise.For solving the limitations of existing DGA-based fault diagnosis method and improve accuracy of fault diagnosis further,a kind of fault diagnosis method for the oil immersed power transformer based on the capsule network(CapsNet)is proposed.The proposed method,by combining the unique capability of CapsNet to handle vectors and the characteristics of dissolved gases in transformer oil,can be able to accurately model the intricate nonlinear relationship between dissolved gases and types of faults,realizing the automatic extraction of key features of the realized training of the model through dynamic routing technique and back propagation algorithm.The test result based on real DGA data shows that the proposed method achieves 90.48%、82.46%and 87.93%on the fault diagnosis accuracy,macro average and recall geometric mean,and it outperforms both SVM and DBN,demonstrating its advantages in real-world transformer fault diagnosis applications.

关 键 词：电力变压器 DGA 故障诊断 胶囊神经网络 动态路由算法 

分 类 号：TM41[电气工程—电器] TP183[自动化与计算机技术—控制理论与控制工程]