考虑DGA数据不平衡性的电力变压器IACGAN故障诊断方法

Transformer fault diagnosis method based on IACGAN under incomplete and unbalanced data sets

不同类别DGA数据之间的不平衡性,使得电力变压器故障诊断模型偏好于将待测样本识别为训练样本中占比更高的类,致使故障诊断精度变低。鉴于此问题,提出一种改进型辅助分类生成对抗网络(Improved auxiliary classifier gener-ative adversarial network,IACGAN)模型用于电力变压器数据生成与故障诊断。首先,构建适用于电力变压器DGA数据结构的辅助分类生成对抗网络(Auxiliary classifier generative adversarial network,ACGAN);同时,构建基于跳跃连结结构的深度残差网络(Deep residual network,ResNet)用作ACGAN的判别器、生成器与辅助分类器,以防止模型加深时的退化问题,以上共同构成IACGAN方法。其次,提取出训练好的IACGAN中的判别器与辅助分类器,得到精确的故障诊断模型。实例分析表明,相比于传统方法SMOTE过采样与ACGAN,所提方法故障诊断性能具有明显提升。

The insufficiency and imbalance between different types of data make the fault diagnosis model prefer to recognize the sample to be tested as the class with a higher proportion of the training sample,which is an important reason for the low fault diagnosis accuracy.To solve this problem,proposes an improved auxiliary classifier generative adversarial network(IACGAN)model for power transformer data generation and fault diagnosis.First,construct an auxiliary classifier generative adversarial network(ACGAN)suit-able for the DGA data structure of power transformers;at the same time,construct a deep residual network(ResNet)based on the shortcut connection structure as discriminator,generator and auxiliary classifier,to prevent the degradation problem when the model is deepened,the above together constitute the IACGAN method.Secondly,extract the discriminator and auxiliary classifier in the trained IACGAN to obtain an accurate fault diagnosis model.The experimental results show that compared with the traditional methods SMOTE oversampling and ACGAN,the proposed method has a significant improvement in fault diagnosis performance.