基于多粒度注意力机制的隔离开关故障诊断  

Disconnector fault diagnosis based on multi-granularity attention mechanism

作　　者：解骞 刘柏泽 丁进中 闫大鹏 杨晓萍[1] 党建[1] XIE Qian;LIU Baize;DING Jinzhong;YAN Dapeng;YANG Xiaoping;DANG Jian(School of Electrical Engineering,Xi’an University of Technology,Xi’an 710054,China;State Grid Urumqi Electric Power Supply Company,Urumqi 830054,China;School of Electrical Engineering,Xi’an Jiaotong University,Xi’an 710049,China)

机构地区：[1]西安理工大学电气工程学院,陕西西安710054 [2]国网乌鲁木齐供电公司,新疆乌鲁木齐830054 [3]西安交通大学电气工程学院,陕西西安710049

出　　处：《电工电能新技术》2024年第10期71-84,共14页Advanced Technology of Electrical Engineering and Energy

基　　金：国家自然科学基金项目(52009106)。

摘　　要：针对现有的大多数深度学习方法只能在有限的含标签样本数据下工作,使诊断模型过拟合,导致模型训练时准确率高而投入运用时故障识别准确率低的问题,本文研究隔离开关在不同工况小样本数据集的准确率高诊断方法,构造应用于不同工况下隔离开关故障诊断的多粒度注意力机制(MG-AM)网络框架。此框架首先要对所获得的隔离开关故障数据进行数据预处理,在此程中将获得增强的数据样本以及数据特征库。随之利用时间对比模块对数据故障进行粗比对,初步获取故障工况的几种可能;并通过多粒度语境对比模块对原始数据预测及预测结果与增强数据进行比对。其次充分挖掘并应用已经搜集的样本资源,以含标签和无标签为输入,网络通过半监督以及无监督学习进行优化,以强化输入数据的处理效果。最终搭建诊断模型,实现对未知样本的故障识别。实验结果表明,所设计的网络可以有效利用固有样本对进行故障识别,对目标的平均识别率达到96.47%。In view of the problem that most existing deep learning methods can only work with limited labeled sample data,which makes the diagnosis model too serious,resulting in high accuracy when training the model but low fault identification accuracy when put into use,this paper studies isolation switches a high-accuracy diagnosis method for small data sample sets in different working conditions,and a Multi-Granular Attention Mechanism(MG-AM)network framework for checking isolation switch fault diagnosis under different working conditions is constructed.First,this framework preprocesses the isolation switch fault data to obtain enhanced data samples and data feature libraries.Next,the time comparison module is used to compare the fault data roughly,and several possibilities of the fault condition are preliminarily obtained.The original data are predicted by the multi-granularity context comparison module,and the predicted results are compared with the enhanced data.Then,making full use of the collected sample data,the labeled and unlabeled sample data are input into the network,and the network is optimized simultaneously through semi-supervised learning and unsupervised learning.Finally,the isolation switch fault diagnosis model is established to realize the accurate identification of the unknown sample fault data.The experimental results show that the MG-AM network framework can effectively use the inherent samples for fault diagnosis,and has a good recognition rate,with an average recognition rate of 96.47%.

关 键 词：隔离开关 故障诊断 对比学习 多粒度注意力机制(MG-AM) 半监督学习 

分 类 号：TM643[电气工程—电力系统及自动化]