DNN在分布式光纤铁路轨道病害检测系统中的应用  

Application of DNN in Distributed Optical Fiber Railway Track Defect Detection System

作　　者：刘康 董武 谢浪 刘旭 贺天才 赵讯 LIU Kang;DONG Wu;XIE Lang;LIU Xu;HE Tiancai;ZHAO Xun(Power Dispatching Control Center of Guizhou Power Grid Co.,Ltd.,Guiyang 550002,China;Key Laboratory of Optical Fiber Sensing and Communication Education Ministry,University of Electronic Science and Technology of China,Chengdu 611731,China)

机构地区：[1]贵州电网电力调度控制中心,贵阳550002 [2]电子科技大学光纤传感与通信教育部重点实验室,成都611731

出　　处：《传感器世界》2023年第11期21-27,共7页Sensor World

摘　　要：针对现有轨道病害识别方法精度较低的问题,提出了一种基于深度学习的轨道病害识别方法。在基于CNN模型结构的VGG和ResNet两种深度学习模型中构建环境,并针对模型优化方法及硬件平台,分别构建1D-CNN、1D-VGG和1D-ResNet模型。对比研究3种模型的识别效果与模型性能,并构建铁路轨道病害信号的频谱数据集用作模型的训练测试。实验结果表明,1D-CNN3、1D-VGG16_bn和1D-ResNet34分别为对应种类模型的最优模型,召回率分别为93.0%、96.1%和96.3%,识别速度为1.20 ms、4.45 ms和8.62 ms,最后综合对比模型的识别性能指标与识别速度,结果表明,1D-VGG16_bn模型具有更快的识别速度,可以更好地应用于基于DAS系统的铁路轨道病害实时在线检测。In the study,a deep learning-based approach is proposed to address the issue of low accuracy in existing railway track defect recognition methods.Two deep learning models,VGG and ResNet,based on the Convolutional Neural Network(CNN)architecture,are constructed to build an environment suitable for railway track defect recognition.To optimize the models and select the appropriate hardware platform,three models,namely 1D-CNN,1D-VGG,and 1D-ResNet,are developed.A dataset of spectral data from railway track defect signals is created for training and testing the models.Comparative analysis is performed on the recognition performance and model efficiency of the three models.Experimental results demonstrate that 1D-CNN3,1D-VGG16_bn,and 1D-ResNet34 are the optimal models for their respective categories,with recall rates of 93.0%,96.1%,and 96.3%.The recognition speeds for these models are 1.20 ms,4.45 ms and 8.62 ms,respectively.Finally,considering both the recognition performance metrics and speed,it is determined that the 1D-VGG16_bn model exhibits faster recognition speed and is better suited for real-time online detection of railway track defect using distributed acoustic sensing(DAS)systems.

关 键 词：深度学习 分布式光纤传感 特征提取 轨道病害检测 

分 类 号：TN212.6[电子电信—物理电子学]