为实现变压器铁心松动故障的识别，提出基于变压器声纹的ResNet卷积神经网络（CNN）用作铁心松动故障的识别，分别比较了相同卷积神经网络ResNet在交叉熵损失函数（SE-ResNet-Dense）和特征表达的角度空间中最大化分类界限的加性角度裕度损失函数(SE-ResNet-ArcLoss)不同表现效果。通过变压器空载试验采集变压器铁心在额定预紧力、松动20%、松动40%时的噪声信号，通过离散傅里叶变换将采集的声纹信号生成时频矩阵，并使用Mel滤波器对时频矩阵降维生成尺寸大大缩小的Mel-语谱图。将采集的噪声信号制作成数据集后输入到两种模型中训练，最终测试集在模型SE-ResNet-Dense上的预测结果为90.753%，在模型SE-ResNet-ArcLoss上的预测结果为97.541%。结果验证SE-ResNet-ArcLoss最适用于变压器铁心松动故障识别。

In order to realize the identification of transformer core looseness faultResNet convolution neural network (CNN) based on transformer voiceprint is proposed as the identification of core looseness fault. The different performance effects of the same ResNet-CNN in the cross entropy loss function (SE-ResNet-Dense) and the additive angular margin loss function (SE-ResNet-ArcLoss) which maximizes the classification boundary in the angle space of feature expression are comparedrespectively. The noise signals of the transformer core at the rated preload20% looseness and 40% looseness are collected through the transformer no-load experiment. The collected voiceprint signal is generated into the time-frequency matrix through the discrete Fourier transformand the Mel filter is used to reduce the dimension of the time-frequency matrix to generate the Mel spectrogram with greatly reduced size. The collected noise signal is made into a data set and input into the two models for training. The prediction result of the final test set on the model SE-ResNet-Dense is 90.753%and that on the model SE-ResNet-ArcLoss is 97.541%. The results show that SE-ResNet-ArcLoss are most suitable for transformer core looseness identification.

TM41