基于通道注意力与残差卷积神经网络的变压器故障诊断

Fault diagnosis of transformer based on Squeeze-and-Excitation mechanism and residual convolutional neural network

为了提高变压器故障诊断的准确性,提出一种基于通道注意力与残差卷积神经网络的变压器故障诊断方法。在卷积神经网络中,考虑不同通道信息间的差异,引入通道注意力,自适应调整不同卷积通道的权重;为了尽可能保留网络层间的差异信息,在网络中引入了残差网络;同时,采用跨网络层的连接方式进一步融合差异信息,充分挖掘油中溶解气体信息和变压器运行状态间的内在关系。在相同数据集下,与不含通道注意力的模型进行对比分析。试验结果表明,所提出的诊断方法具有更好的诊断准确性和稳定性,其诊断平均准确率达到了95.07%。

In order to improve the accuracy of transformer fault diagnosis,a transformer fault diagnosis method based on Squeeze-and-Excitation mechanism and residual convolutional neural network.In the convolutional neural network,considering the difference of different channel information,Squeeze-and-Excitation mechanism is introduced to adaptively adjust the weight of different convolution channels.In order to preserve the difference information between network layers as much as possible,the residual network is introduced in the network.The cross-network layer connection method is used to further integrate the difference information,and fully explore the relationship between the dissolved gas information in the oil and the operating state of the transformer.Under the same data set,a comparative analysis was performed with the model without channel attention.The experimental results show that the diagnostic method proposed has better diagnostic accuracy and stability,and its average diagnostic accuracy rate reaches 95.07%.