http://www.jsjkx.comDOI:10.11896/jsjkx.220600238到稿日期:2022-06-27返修日期:2022-10-18通信作者:雷印杰(yinjie@scu.edu.cn)基于细粒度星座图识别的光性能监测方法陈进杰贺超肖枭雷印杰四川大学电子信息学院成都610065(chenjinjie@scuailab.com)摘要在光纤通信中,传统光性能监测(OpticalPerformanceMonitoring,OPM)主要依靠分析信号的时频域信息来实现,但此类方法无法完成多任务联合监测,因此其灵活性较低。随着机器学习的发展,基于机器学习的光信号调制格式(Modulation-Format,MF)及光信噪比(OpticalSignalNoiseRatio,OSNR)监测方法被逐渐应用。但现有方法未考虑信号的细粒度特征,因此在复杂场景中对OSNR的监测精度较低。针对上述问题,文中提出了一种基于细粒度星座图识别的光信号MF和OSNR联合监测模型(Fine-GrainedOpticalPerformanceMonitorNetwork,FGNet)。首先,在骨干特征提取模块中采用深度残差结构对星座图进行深度特征提取;其次,提出多层双线性池化(MultilayerBilinearPooling)模块,对星座图特征进行细粒度特征分析;最后,提出联合监测模块对MF和OSNR进行特征融合分析。在拥有7200张星座图的仿真数据集中进行广泛的实验,实验结果表明,所提方法相比现有方法取得了更优越的性能。关键词:机器学习;光信噪比监测;调制格式分类;细粒度图像识别;残差神经网络中图法分类号TP389OpticalPerformanceMonitoringMethodBasedonFine-grainedConstellationDiagramRecognitionCHENJinjie,HEChao,XIAOXiaoandLEIYinjieCollegeofElectronicsandInformationEngineering,SichuanUniversity,Chengdu610065,ChinaAbstractInopticfibercommunication,traditionalopticalperformancemonitoring(OPM)mainlyreliesonanalyzingthetime-fre-quencydomaininformationofthesignal.However,conventionalmethodscannotcompletemulti-taskjointmonitoring,sotheyarelessflexible.Withthedevelopmentofmachinelearning,themonitoringofopticalsignalmodulationformat(MF)andopticalsig-nal-to-noiseratio(OSNR)basedonmachinelearninghavebeengraduallyapplied.However,existingmethodshavelowaccuracyforOSNRmonitoringincomplexscenariosbecausetheydonotconsiderthefine-grainedcharacteristicsofthesignal.Thispaperproposesajointmonitoringmodel(FGNet)foropticalsignalMFandOSNRbasedonfine-grainedconstellationidentificationtosolvethisproblem.Firstly,theback...
