基于残差注意力自适应去噪网络和Stacking集成学习的局部放电故障诊断  

Partial discharge fault diagnosis based on residual attention adaptive denoising network and stacking ensemble learning

作　　者：廖晓青 陈历 许建远 金宝权 姜自超 刘俊峰[2] Liao Xiaoqing;Chen Li;Xu Jianyuan;Jin Baoquan;Jiang Zichao;Liu Junfeng(Maoming Power Supply Bureau of Guangdong Power Grid Co.,Ltd.,Maoming 525000,China;School of Automation Science and Engineering,South China University of Technology,Guangzhou 510641,China)

机构地区：[1]广东电网有限责任公司茂名供电局,广东茂名525000 [2]华南理工大学自动化科学与工程学院,广东广州510641

出　　处：《电子技术应用》2024年第11期66-73,共8页Application of Electronic Technique

基　　金：南方电网公司科技项目(030900KC23040006)。

摘　　要：针对传统局部放电(Partial Discharge,PD)故障诊断方法在处理复杂含噪PD信号存在局限性并依赖于人工去噪和专家经验,难以学习到PD特征多样化表达等问题,分别提出残差注意力自适应去噪网络(Residual Attention Adaptive Denoising Network,RAADNet)和基于Stacking集成学习的PD故障诊断模型。RAADNet基于残差网络结构设计,通过集成CAM注意力机制和软阈值函数实现自适应去噪;Stacking集成模型的基学习器分别由基于卷积神经网络的RAADNet、基于多头自注意力机制的Transformer以及基于Boosting集成策略的XGBoost多个差异化模型共同构建构成。实验结果表明,提出的RAADNet优于其他先进方法,识别准确率达到93.99%,Stacking集成模型则通过学习多样化特征表达,进一步提高模型性能,达到96.79%识别准确率。To overcome the challenges posed by traditional Partial Discharge(PD)fault diagnosis methods,such as their inability to effectively process complex,noisy PD signals and their reliance on manual denoising and expert input,and the difficulty on learning diverse PD feature representations,this paper introduces two advanced solutions:Residual Attention Adaptive Denoising Network(RAADNet)and Stacking ensemble-based PD fault diagnosis model.RAADNet leverages a residual network framework integrated with a Channel Attention Module(CAM)and a soft thresholding function for adaptive noise reduction.The Stacking ensemble model comprises distinct base learners,including the RAADNet with convolutional neural network architecture,a Trans‐former featuring multi-head self-attention,and an XGBoost model that adopts a Boosting strategy.Experimental findings reveal that RAADNet surpasses competing advanced techniques,achieving an accuracy of 93.99%.The Stacking model further improves performance by leveraging diverse feature representations,reaching an accuracy of 96.79%.

关 键 词：气体绝缘开关柜 局部放电 Stacking集成学习 卷积神经网络 TRANSFORMER 

分 类 号：TM85[电气工程—高电压与绝缘技术] TN91[电子电信—通信与信息系统]