结合注意力机制和多尺度特征融合的变压器局部放电诊断  

Partial Discharge Diagnosis of Transformer Combining Attention Mechanism and Multi-scale Feature Fusion

作　　者：张玥 朱永利[1] 钱涛 ZHANG Yue;ZHU Yongli;QIAN Tao(Department of Computer,North China Electric Power University,Baoding 071003,China;Department of Electrical Engineering,North China Electric Power University,Baoding 071003,China)

机构地区：[1]华北电力大学计算机系,河北保定071003 [2]华北电力大学电力工程系,河北保定071003

出　　处：《电力科学与工程》2025年第4期43-51,共9页Electric Power Science and Engineering

基　　金：河北省自然科学基金资助项目(F2022502002)。

摘　　要：针对深度卷积网络对变压器局部放电边际谱特征提取能力不足的问题,提出了一种结合注意力机制和多尺度特征融合两模块的变压器局部放电识别方法。首先,针对局部放电领域现有时频分析方法的不足,提出采用逐次变分模态分解对局部放电信号进行分解,再用Hilbert变换来获得边际谱,作为网络模型的输入;然后,将ResNet作为基础模型,在残差块中引入坐标注意力模块,以提高模型对局放信号边际谱中重要区域的关注度;最后,搭建多尺度特征融合模块对网络各阶段所提取的特征进行融合,使模型能在提取深层语义特征的同时保留浅层中提取到的局部细节信息,以增强模型的表征能力。仿真实验结果表明:该方法能有效实现变压器局部放电类型诊断,达到96.5%的识别准确率,超过其他经典深度卷积诊断模型。Aiming at the problem that the deep convolutional network has insufficient ability to extract the marginal spectrum feature of transformer partial discharge,a transformer partial discharge recognition method combining attention mechanism and multi-scale feature fusion is proposed.Firstly,in view of the shortcomings of the existing time-frequency analysis methods in the field of partial discharge,successive variational mode decomposition(SVMD)is proposed to decompose the partial discharge signal,and then Hilbert transform is used to obtain the marginal spectrum as the input of the network model.Secondly,the ResNet is used as the basic model,and the coordinate attention(CA)module is introduced into the residual block to improve the model’s attention to the important regions in the marginal spectrum of the local amplification signal.Finally,a multi-scale feature fusion module is built to fuse the features extracted at each stage of the network,so that the model can extract the deep semantic features while retaining the local detail information extracted from the shallow layer,and enhance the representation ability of the model.The simulation results show that the proposed method can effectively realize the partial discharge type diagnosis of transformers,and achieve 96.5%recognition accuracy,which is higher than other classical deep convolution diagnosis models.

关 键 词：变压器 局部放电 卷积神经网络 Hilbert边际谱 注意力机制 特征融合 

分 类 号：TM85[电气工程—高电压与绝缘技术]