基于结构优化卷积深度网络的变压器溶解气体分析

Transformer Dissolved Gas Analysis Based on Structure-Optimized Convolutional Deep Network

针对变压器故障诊断准确率低下的问题,提出一种结构优化卷积长短期记忆网络的变压器故障诊断模型。采用卷积神经网络提取隐含数据的空间特征,再由长短期记忆神经网络提取隐含数据的时间特征,在训练CNN的超参数时,网络的结构参数由GA更新计算,这能够有效地防止训练模型出现过拟合现象,同时也解决了训练过程陷入局部最优的问题。并基于油中溶解气体分析(DGA)技术,对变压器故障数据进行预处理,使其能作为训练模型的输入,Softmax函数被用作整个网络的输出层,以确定故障的类型。同时采用ROC和PRC作为模型训练性能的评价标准,结果表明,所提出的模型诊断准确率高于CNN、LSTM、CNN-LSTM、GAC-NN故障诊断模型准确率,验证了所提方法能够有效地提升变压器故障诊断性能。

In view of the low accuracy of transformer fault diagnosis,a structure-optimized convolutional long short-term memory network of transformer fault diagnosis model is proposed.Convolution neural network is used to extract the space characteristics of the hidden data,and then long short-term memory network is adopted to extract the time characteristics of the hidden data.In the training of CNN’s super parameters,the structure parameters of the network are calculated by the GA updates,which can effectively prevent the occurrence of the over-fitting of the training model.At the same time,it solves the problem that the training process falls into local optimum.Based on dissolved gas analysis(DGA)technology,the transformer fault data are preprocessed,so that they can be used as the model input for network training,and the output layer is used to get the fault diagnosis type by using Softmax function.At the same time,ROC and PRC are used as the evaluation criteria for the model’s training performance.The results show that the diagnostic accuracy of the proposed model is higher than that of the CNN,LSTM,CNN-LSTM and GA-CNN fault diagnosis models,which verifies that the proposed method can effectively improve the performance of transformer fault diagnosis.