基于人工智能的电信网络的思考（英）

ConsiderationofAIbasedTelecomNetwork ZhangHao ChinaMobile 2024.3 Threerevolutionstagesofcommunicationnetworks NetworkIT-lised NetworkAI-lised revolution Growingdemandforbusinessdiversityandresilience ITrapidlychangingtheshapeofthenetwork. NetworkIP-lisedrevolution revolution 2000s 2010s 2020s 2030s ExplosivegrowthinInternetbandwidthandsubscribernumbers.Transformationfromsinglevoiceservicetoconvergedservice. Narrowbanddigitalvoicelow-speeddata Businessdemand Fromcircuitswitchingtopacketswitching FromswitchingIPtoend-to-endfullIP Technicalfeatures High-speeddatainbroadbanddigitalvoice Peopletopeople--Peopletothings--ThingstothingsCommunicationsInternetserviceprovider Intelligence,universalconnectivity,elementalconvergence SDNNFVSBA IP、MPLS、SRv6PTN Softswitch、IMS、VoLTE RoCE、NVLink、GSEMLOps FederalLearning NetworkforAI AIforNetwork、AINative FromClosedNetworkstoOpenBusiness FromRigiditytoResilience AINeedsNetworksforEfficientCommunicationPerformance. NetworksneedAItoenhancehigher-orderself-intelligencecapabilities Typicaltechnology Communicationnetworkispromptingthenewinformationtechnologytoallareasofdeeppenetration 3 Morecomprehensivenetworkperformancemeasures Networkcomplexityincreasesfromgenerationtogeneration AIisthekeypathtomeetthenewmetricsofmobilecommunicationnetworks,empoweringnetworkstoimprovenetworkoperationefficiency. AIfornetwork Operationsandmaintenance Majorrequirementsandchallengesofnetworkintelligence Networksneedtobequicklyadaptedtothecustomisedrequirementsofdiversescenarios Theeraofuniversalintelligencerequiresefficientcommunicationperformanceofnetworks. Majorshiftfrommobilecommunicationstomobileinformationservices IntelligenceInclusiveUbiquitous Networkintelligencegeneratednatively AINative AIgeneratednatively The6Gmobileinformationnetworkwillprovidethewholeprocessofinformationflowservices,achievingthebasicplatformofAIubiquityanduniversality. Intelligentcomputinginfrastructure’senhancement WideAreaServiceUniversality NetworkforAI StrengtheningComputibilitywithNetworks DistributedtrainingofGPUclustersbringsalargeamountofcommunicationoverhead,andnetworkperformancebecomesabottleneckrestrictingAIarithmeticenhancement TheconvergenceofnetworkandAIincludestwoaspects:"networkenablesAI"and"AIempowersnetwork".ThefirstistoprovideanuniversalaccesstoAIservices.ThesecondisAI-enablednetworkstoimprovenetworkoperationandO&Mefficiency. AIforNetwork:Twomajorscenarios Byextractingtheregularfeaturesofdataincomplexscenarios, AIcanhelpthetelecomnetworkimproveefficiencyinmaintenanceandoperationscenarios Aroundtheentirelifecycleofnetworkplanning,construction,maintenance,optimization,andoperation,AIoptimizestheprocesstoachievethecostreductionandefficiencyenhancement AIreconstructsthenetworkoperationprocesstoachievethebestmatchingofnetworkresources,operatingefficiencyanduserexperience NetworkMaintenanceNetworkOperation AIreplacinghuman Richscenarios Centralizeddata AIreplacingequipment Simplescenario Discretedata Offlineanalyzing Maintenance+AI Unifieddatainterface Onlineprocessing Operation+AI Poordatastandardization Lowercost Lowerpowerconsumption LowercostLowerpowerconsumption AIforNetwork:Intelligentnetworkmaintenance AsglobaloperatorscontinuetoevolvetowardsL4orhigher-levelAutonomousNetworks(ANs),networkmaintenancemodeupgradesfromautomationtointelligence.AnintelligentsystemwithAImodelsisbecomingthetrendoffuturetelecomnetworkmaintenance. FundamentalNetworkModels&Tools SpecializedModels (Alarm/Log/Perf/…) AutomationTools (SI/Config/Visualize) Theintelligentnetworkmaintenancesystem Visualization FaultDiagnosis •VisualizationofFCAPSdata ICTLargeLanguageModel ModelOptimizationforICTScenario ICTPrivateDomainKnowledgeData LargeLanguageModel(LLM)Base •Visualizationofnetworktopology •Hiddendangerprediction •Intelligentrootcaseanalysis IntelligentGuidance KeepLive&Recovery •SmartQ&A •Onboardingguidance •Disasterrecoveryassessment •Automaticexecution RoutineMaintenance NetworkOptimization ProcedureIntegration •In-depthinspection •Healthassessment •QualityImprovement •Energyefficiencyoptimization AIforNetwork:Intelligentnetworkoperation NWDAF(NetworkDataAnalyticsFunction)isintroducedtorealizeoptimalmatchofnetworkresource,achievingthehighestnetworkefficiencyandbetteruserexperience. ArchitectureTypicalCase AI UDRNEFNWDAFAF NudrNnefNnwdafNaf PCFNSSF NpcfNssf ADRF Nadrf Noam Namf Nsmf Nupf Ndccf Nmfaf OAM AMF SMF N4 DCCf MFAF UE RAN UPF Initiation Recognition Inspection Perception&Identification Observing ApplicationtypeServiceexperiencedata RANloadstatus Calculating KPIstatisticsSneaker QoEexperience ...... Controling QOSAssuranceSneakerMarketing ...... Visualizing RegionalLevelApplicationLevelUELevel ...... NWDAFbasednetworkoperationintelligence •ServiceRegistration:Servicearea,AnalyticsID. •Datacollection:5GCNF、AF、OAM,real-timecustomizedcollection. •AI/MLtraining:Providemachinelearningmodels. •Inferenceperforming：Prediction/statistics/recommendation. •AnalyticsFeed