基于注意力网络的小样本高速铁路接触网异常检测  

Few-Shot Anomaly Detection for High-Speed Railway Catenary Based on Attention Networks

作　　者：秦加辉 陈圣堂 吴泽彬[1] 徐洋 詹天明[3] 章来胜 丁道华 舒青青 王静怡 Qin Jiahui;Chen Shengtang;Wu Zebin;Xu Yang;Zhan Tianming;Zhang Laisheng;Ding Daohua;Shu Qingqing;Wang Jingyi(School of Computer Science and Engineering,Nanjing University of Science and Technology,Jiangsu,Nanjing 210094,China;Nanjing Power Supply Section,China Railway Shanghai Group Co.,Ltd.,Jiangsu,Nanjing 210011,China;School of Computer Science,Nanjing Audit University,Jiangsu,Nanjing 211815,China)

机构地区：[1]南京理工大学计算机科学与工程学院,江苏南京210094 [2]中国铁路上海局集团有限公司南京供电段,江苏南京210011 [3]南京审计大学计算机学院,江苏南京211815

出　　处：《铁道技术标准(中英文)》2025年第1期47-53,60,共8页Railway Technical Standard(Chinese & English)

摘　　要：针对高速铁路接触网异常检测的问题,提出基于注意力网络的小样本高速铁路接触网异常检测方法。该方法能够在有限的缺陷标注数据情况下,通过有效利用大规模非高速铁路领域数据来提升模型的泛化能力,并通过接触网数据进行迁移学习从而达到接触网异常检测的目的。在实际检测过程中,证明了小样本检测方法的可行性和有效性。这一方法将为高速铁路接触网异常检测提供强大支撑,在缺陷样本稀少的情况下,提高接触网异常检测的精度和效率。This paper addresses the anomaly detection in high-speed railway catenary systems by proposing a fewshot learning method based on attention networks.The proposed method enhances model generalization through the effective utilization of large-scale data from non-high-speed railway domains under conditions of limited annotated defect samples.It achieves the goal of detecting anomalies in catenary systems via transfer learning using catenaryspecific data.Experiments conducted during the actual detection process demonstrate the feasibility and effectiveness of this few-shot detection approach.This method provides robust support for high-speed railway catenary anomaly detection,significantly improving accuracy and efficiency even when defect samples are scarce.Such improvements are crucial to ensuring the safe and rapid operation of high-speed railways.

关 键 词：注意力网络 小样本学习 接触网异常检测 迁移学习 

分 类 号：U225.4[交通运输工程—道路与铁道工程]