基于轻量级卷积神经网络的X-DR接地故障检测  

X-DR Earthing Fault Inspection Based on Lightweight Convolution Neural Network

作　　者：崔玉璟 程宏波 CUI Yujing;CHENG Hongbo

机构地区：[1]中铁电气化铁路运营管理有限公司 [2]华东交通大学电气与自动化工程学院

出　　处：《电气化铁道》2025年第2期11-15,共5页Electric Railway

摘　　要：气体绝缘全封闭组合电器(GIS)故障检测目前多依赖于人工检查和定期维护,存在效率低、准确性差、易漏检等问题。本文提出一种基于轻量级网络ShuffleNetV2的改进模型,结合X射线数字成像技术(X-DR)自动识别GIS设备接地故障。以ShuffleNetV2网络模型为基础,采用新的激活函数FReLU,可以捕获空间上下文信息,促进网络学习;引入CBAM注意力机制提高网络对细化特征的提取能力,提升分类网络对X-DR图像识别的适应性;通过迁移学习初始化主干网络权重,增强模型的训练起点,缩短训练时间,同时解决训练数据集不足的问题。试验结果表明,改进后的ShuffleNetV2在自建数据集上的平均准确率为91.2%,模型参数量仅为1.61 M,与已有的经典模型相比,分类性能更佳,为电气化铁路牵引供电设备故障检测提供了新思路。The inspection of faults of full-enclosed gas insulated switchgear(GIS)currently relies heavily on manual inspection and regular maintenance,and there exists problems such as low efficiency,poor accuracy,and missing inspection.The paper puts forward an improved model based on the lightweight network ShuffleNetV2,which combines X-ray digital imaging technology(X-DR)to automatically identify the earthing faults in GIS equipment.Based on the ShuffleNetV2 network model,a new activation function FReLU is adopted to capture spatial context information and promote network learning.CBAM attention mechanism is introduced to enhance the network's ability to extract refined features and improve the classification network's adaptability to X-DR image recognition.The weights of the backbone network is initialized through transfer learning to enhance the training starting point of the model,shorten the training time and solve the problem of insufficient training data set.The experimental results show that the improved ShuffleNetV2 has an average accuracy of 91.2%on a self built data set,with a model parameter size of only 1.61 M.Compared with existing classical models,it has better classification performance and provides a new idea for inspection of faults of traction power supply equipment in electrified railways.

关 键 词：接地故障 X射线 ShuffleNetV2 注意力机制 迁移学习 

分 类 号：U226.7[交通运输工程—道路与铁道工程]