基于相似度机制AdaBoost-DBN的变压器故障层级诊断  

Fault Hierarchical Diagnosis of Transformer Based on AdaBoost-DBN Similarity Mechanism

作　　者：徐耀松[1] 李佳旺 段彦强 XU Yaosong;LI Jiawang;DUAN Yanqiang(Faculty of Electrical and Engineering Control,Liaoning Technical University,Liaoning Huludao 125105,China)

机构地区：[1]辽宁工程技术大学电气与控制工程学院,辽宁葫芦岛125105

出　　处：《高压电器》2023年第6期154-164,共11页High Voltage Apparatus

基　　金：国家自然科学基金(51974151);辽宁省教育厅科技项目(LJ2019QL015)。

摘　　要：针对传统的变压器故障诊断方法诊断精度以及泛化能力有限,提出了一种基于相似度机制AdaBoost-DBN的变压器故障层级诊断方法。首先应用层级分类思想把变压器故障诊断模型分为初级分类器和次级分类器;然后初级分类器采用Bagging-XGBoost方法,使其诊断结果与变压器油中溶解气体数据进行特征融合作为次级分类器的输入特征;最后次级分类器采用相似度机制AdaBoost-DBN方法,使用AdaBoost集成学习有效避免DBN模型过拟合效应,通过提出相似度机制为AdaBoost网络的初始样本权重赋值,克服随机因数对诊断结果的影响,进一步提高模型的诊断精度以及泛化能力。实验结果表明,相比于BP神经网络、支持向量机与DBN网络,文中方法的诊断准确率分别提高17.9%、17.3%、16.1%。In view of the limited diagnostic accuracy and generalization ability of fault diagnosis methods of traditional transformer,a kind of fault hierarchy diagnosis method of transformer based on the similarity mechanism AdaBoost-DBN is proposed.Firstly,the transformer fault diagnosis model is divided into primary and secondary classifier by the idea of hierarchy classification.Then,the Bagging-XGBoost method is used for the former,and the diagnosis results are fused with the data of dissolved gas in transformer oil as the input features of the secondary classifier.Finally,the secondary classifier adopts the similarity mechanism AdaBoost-DBN method,and the AdaBoost ensemble learning is used to effectively avoid the over-fitting effect of DBN model.The similarity mechanism is proposed to give the weight of the initial sample of AdaBoost network,so as to overcome the influence of random factors on the diagnostic results and improve the diagnostic accuracy and generalization ability of the model further.The experimental results show that compared with BP neural network,the support vector machine and DBN network,the diagnostic accuracy of the proposed method is improved by 17.9%,17.3%and 16.1%,respectively.

关 键 词：电力变压器 故障诊断 油中溶解气体分析 集成学习 相似度 层级分类 

分 类 号：TM407[电气工程—电器]