文章编号：1005-5630(2023)02-0018-08DOI:10.3969/j.issn.1005-5630.2023.002.003基于轻量级网络的光纤传感振动信号识别陈玲玲，李柏承，张大伟，杨涵，吴春波（上海理工大学光电信息与计算机工程学院，上海200093）摘要：虽然传统卷积神经网络的识别率很高，但是其庞大的参数量会导致工业部署困难，且识别响应速度慢。引入轻量级卷积神经网络MobileNet，使用深度可分离卷积替代传统卷积，大大减少了模型参数量。以MobileNet为基准网络，实现了基于一维轻量级网络MobileNet-18的Φ-OTDR周界入侵事件识别。通过实验对比了不同结构下的网络识别率和识别速度，在保证模型的准确率不会大幅度降低的情况下，选取MobileNet-18作为最佳模型。采集了攀爬、切割、风吹、举起、拉动和走动这6种周界光纤入侵信号。在6种光纤入侵信号识别中，MobileNet-18达到了识别率为98.33%，响应时间为9.27ms的最佳效果关键词：卷积神经网络；轻量级网络；深度可分离卷积；光纤信号；周界安全中图分类号：TN913.7文献标志码：AOpticalfibersensingvibrationsignalrecognitionbasedonlightweightnetworkCHENLingling，LIBaicheng，ZHANGDawei，YANGHan，WUChunbo（SchoolofOptical-ElectricalandComputerEngineering,UniversityofShanghaiforScienceandTechnology,Shanghai200093,China）Abstract:Basedontheapplicationofdistributedopticalfibersensingsysteminthefieldofperimetersecuritymonitoring,thereareproblemssuchasslowresponsespeedandlowrecognitionrate.Althoughtherecognitionrateofthetraditionalconvolutionalneuralnetworkisveryhigh,itshugeamountofparametersmakesindustrialdeploymentdifficultandtherecognitionresponsespeedisslow.ThispaperintroducesthelightweightconvolutionalneuralnetworkMobileNet,whichusesdepth-separableconvolutiontoreplacethetraditionalconvolution,whichgreatlyreducestheamountofmodelparameters.ThispaperusesMobileNetasthebenchmarknetworktoimplementaone-dimensionallightweightnetworkbasedonMobileNet-18Φ-OTDRperimeterintrusioneventrecognition,comparedthenetworkrecognitionrateandrecognitionspeedunderdifferentstructuresthroughexperiments,andselectedMobileNet-18asthebestmodelundertheconditionthattheaccuracyofthemodelwouldnotbegreatlyreduced.Intheexperiment,six收稿日期：2022-12-02基...
