摘要
Φ-OTDR分布式光纤传感系统在安全监测领域应用广泛,其关键的任务是振动事件的类型识别。传统的模式识别方法的识别率和鲁棒性都不够理想,而基于深度学习的方法能自发从数据提取特征并完成分类,准确率和适应性都更好。相比二维卷积神经网络(2D-CNN),一维卷积神经网络(1D-CNN)的网络大小和训练速度均更有优势,本文以LeNet-5为基准网络,实现了基于1D-CNN的Φ-OTDR地埋光纤检测振动事件分类,并通过实验法对比分析了不同结构超参数对识别效果的影响,选取最优参数构建LeNet-1D-V网络。实验结果显示,本文构建的LeNet-1D-V在5种类别的地埋光纤振动事件分类中,将分类准确率从92.3%提升至94.6%,为多事件类型的地埋光纤事件分类研究提供了参考依据。
Φ-OTDR distributed optical fiber sensing system is widely used in the field of safety monitoring,and its key task is to identify the type of vibration event.The recognition rate and robustness of traditional pattern recognition methods are not sufficient enough.Deep learning can extract features and realize classification spontaneously,with better accuracy and adaptability.Compared with the two-dimensional convolutional neural network(2 D-CNN),the one-dimensional convolutional neural network(1 D-CNN)has smaller net size and faster training speed.Set LeNet-5 as the baseline network,this paper attained to classify the vibration events forΦ-OTDR with buried sensing fiber which based on 1 D-CNN.In addition,the optimal parameters to construct a LeNet-1 D-V network was selected by comparing the influence of different structural hyperparameters on recognition accuracy.The experiment results show that the LeNet-1 D-V constructed in this paper improves the classification accuracy rate from 92.3%to 94.6%in the classification of 5 types of buried fiber vibration events,which provides a foundation for the classification of multiple event types detected by buried fiber.
作者
罗天林
王砾苑
施羿
LUO Tian-lin;WANG Li-yuan;SHI Yi(Shantou University,School of engineering,Guangdong Provincial Key Laboratory of Digital Signal and Image Processing,Shantou 515063,China;Jiangxi College of Applied Technolgy,Ganzhou 341000,China)
出处
《光电子．激光》
EI
CAS
CSCD
北大核心
2020年第9期955-964,共10页
Journal of Optoelectronics·Laser
基金
国家自然科学青年基金(61801283)
汕头大学科研启动基金(NTF18007)资助项目。
关键词
分布式光纤传感系统
一维卷积神经网络
结构超参数优化
distributed optical fiber sensing system
one-dimensional convolutional neural network
structural hyperparameter optimization
