引用本文: | 孙 抗,张 浩,杨 林,等.声信号下基于双通道特征融合网络的电抗器故障诊断方法[J].电力系统保护与控制,2025,53(1):104-113.[点击复制] |
SUN Kang,ZHANG Hao,YANG Lin,et al.Fault diagnosis method of reactor based on a dual-channel feature fusion network with an acoustic signal[J].Power System Protection and Control,2025,53(1):104-113[点击复制] |
|
摘要: |
目前,干式电抗器故障诊断方法主要围绕在基于振动信号的机械故障展开,故障类型单一,并且存在传感器安装困难等问题。为此,搭建了基于声信号下的干式电抗器故障试验平台,设置了多种故障类型。为了提升小样本下故障识别的准确率,提出一种基于双通道特征融合网络的干式电抗器故障诊断方法。首先,采用格拉姆角场(Gramian angle field, GAF)进行编码,将一维时序转化为二维图像。其次,采用双通道并行的CNN-ResNet网络结构,引入高效通道注意力机制(efficient channel attention, ECA)来获取二维关键信息,将二维图像特征与一维时序特征进行深度提取与融合。最后,基于有限元仿真来获取源域数据,采用迁移学习方法来获取目标域最优网络参数。试验对比表明:所提方法相比其他方法有着较强的特征提取能力,能够将故障特征显著分离,在小样本下的故障识别准确率最高可达99.5%,同时具有良好的泛化性和收敛速度。 |
关键词: 电抗器 声信号 故障 特征 |
DOI:10.19783/j.cnki.pspc.240363 |
投稿时间:2024-03-20修订日期:2024-05-21 |
基金项目:国家自然科学基金项目资助(U1804143);河南省科技攻关计划项目资助(202102210092) |
|
Fault diagnosis method of reactor based on a dual-channel feature fusion network with an acoustic signal |
SUN Kang1,ZHANG Hao1,YANG Lin2,CHANG Liang2,YANG Ming1 |
(1. School of Electrical and Automation, Henan Polytechnic University, Jiaozuo 454003, China;
2. Jiaozuo Guangyuan Electric Power Group Co., Ltd., Jiaozuo 454150, China) |
Abstract: |
At present, the fault diagnosis methods of a dry reactor mainly focus on mechanical faults based on vibration signals. The fault type is singular, and there are problems such as difficulties with sensor installation. Thus, a dry reactor fault test platform based on an acoustic signal is built, and a variety of fault types are set up. To improve the accuracy of fault identification in small samples, a dry reactor fault diagnosis method based on a dual-channel feature fusion network is proposed. First, the one-dimensional time sequence is converted into two-dimensional image using Gramian angle field (GAF) encoding. Secondly, a two-channel parallel CNN-ResNet network structure is adopted, and an efficient channel attention (ECA) mechanism is introduced to obtain two-dimensional key information. Then the two-dimensional image features and one-dimensional time series features are extracted and fused. Finally, the source domain data is obtained based on finite element simulation, and the optimal network parameters of the target domain are obtained by the transfer learning method. Experimental comparison shows that the proposed method has stronger feature extraction ability than other methods, can significantly separate fault features, and the fault identification accuracy can reach 99.5% with small samples. It has good generalizability and convergence speed. |
Key words: reactor acoustic signal fault features |