考虑样本类内不平衡的CHPOA-DBN变压器故障诊断方法

CHPOA-DBN Transformer Fault Diagnosis Method Considering Sample Within-Class Imbalance

为解决变压器故障样本类内不平衡与人为确定深度信念网络(deep belief network,DBN)的网络参数导致故障诊断精度低的问题,提出一种基于样本均衡和改进DBN的变压器故障诊断方法。首先,针对合成少数类过采样算法(synthesis minority oversampling technique,SMOTE)生成样本加剧类内不平衡的问题,提出基于改进K均值(improved K-means,IK-means)的IK-means SMOTE算法,据此得到类间、类内均衡的故障样本;其次,利用Tent混沌映射改进的鹈鹕优化算法(chaotic hybrid pelican optimization algorithm,CHPOA)对DBN的隐含层节点数、反向微调学习率寻优,构建CHPOA-DBN变压器故障诊断模型;最后,基于实验数据,分别将经典过采样算法、经典故障诊断模型与所提方法进行对比分析,结果表明:所提方法故障诊断准确率达到96.25%,可以为变压器故障样本不均衡条件下的故障智能诊断提供重要参考。

In recent years,deep belief network(DBN)based transformer fault diagnosis methods have been developed.However,they share two prominent drawbacks,which are the low accuracy issue caused by the within-class imbalance of transformer faults samples and the artificial determination of the network parameters of deep belief network(DBN).In this paper,a transformer fault diagnosis method based on sample balance processing and improved DBN is proposed.Firstly,an improved K-means(IK-means)synthesis minority oversampling technique(SMOTE)algorithm is proposed to obtain within-class and between-class balanced fault samples.Then,the Tent chaotic map embedded chaotic hybrid pelican optimization algorithm(CHPOA)is developed to optimize the number of hidden layer nodes and reverse fine-tuning learning rate of DBN,and the CHPOA-DBN transformer fault diagnosis model is constructed.Finally,the classical oversampling algorithm,the classical fault diagnosis model and the proposed method are compared and analyzed,based on the experimental data,respectively.The results show that the fault diagnosis accuracy of the proposed method reaches 96.25%,which provide an important reference for intelligent fault diagnosis under imbalanced fault samples of transformers.