《6GANA：6G網絡原生AI技術需求白皮書（2023）（英文版）（34頁）.pdf》由會員分享，可在線閱讀，更多相關《6GANA：6G網絡原生AI技術需求白皮書（2023）（英文版）（34頁）.pdf（34頁珍藏版）》請在三個皮匠報告上搜索。

1、2022/1/20 1 Abstract The next-generation 6G mobile system aims to support and implement intelligent connection of everything,intelligent association of services and tasks,and inclusive intelligence.It will become a powerful intelligent infrastructure platform that supports and empowers a wide variet

2、y of industries.With the acquisition of consumers at all levels and the in-depth development of digital and intelligent applications in various industries,there is an increasingly urgent need to provide native AI capabilities and services in new 6G networks.This white paper systematically analyzes a

3、nd summarizes the technical requirements and impacts(in terms of capability,service,architecture,standardization,etc.)of native AI design based on the requirements and trends of future 6G business forms,new industry scenarios and use cases,and new technologies to effectively support the design and i

4、mplementation of native AI-related technical solutions for new 6G networks.The organizations that have participated in writing this white paper include but are not limited to:ZTE,CICT Mobile,Huawei,China Mobile,China Telecom,China Unicom,ShanghaiTech University,Chongqing University of Posts and Tele

5、communications,Nokia Shanghai Bell,vivo,Purple Mountain Laboratories,AsiaInfo,Ericsson,Dalian Maritime University,and Peng Cheng Laboratory.Many thanks to all these organizations for their contributions.2 Contents 1 Introduction.4 2 Applications Integrating AI and 5G Mobile Systems.5 2.1 Status Quo

6、of AI for 5G.5 2.2 Status Quo of 5G for AI.5 2.3 Status Quo of 5G Network AI.6 2.4 Challenges in Integrating AI and 5G Mobile Systems.10 3 Mechanism of Native AI in the New 6G Mobile System.11 3.1 Necessity Analysis.11 3.2 Feasibility Analysis.11 3.3 Gain Analysis.12 3.4 Analysis of New Features and

7、 New Paradigms.13 4 Technical Requirements of Native AI in the New 6G Mobile System.15 4.1 Capability Requirements.15 4.1.1 Computing Power Requirements.15 4.1.2 Algorithm Requirements.16 4.1.3 Data Requirements.17 4.1.4 Other Requirements.17 4.2 Service Requirements.18 4.2.1 Computing Service Requi

8、rements.18 4.2.2 Algorithm Service Requirements.19 4.2.3 Data Service Requirements.20 4.2.4 Other Service Requirements.21 4.3 Architecture Requirements.22 4.3.1 Computing Architecture.22 4.3.2 Algorithm Architecture.25 4.3.3 Data Architecture.26 4.3.4 Other Architectures.28 5 Summary of Technical Re

9、quirements and Principles.29 References.30 Acronyms and Abbreviations.31 3 Authors Contributor Organization Li Yang,Feng Xie,Honghui Kang,Yan Xue,Menghan Wang,Jiaohong Niu ZTE Corporation Ming Ai,Xiaoyan Duan,Wanfei Sun,Min Shu CICT Mobile Chenghui Peng,Zhe Liu,Jun Wang,Fei Wang Huawei Technologies

10、Co.,Ltd.Gang Li,Zirui Wen China Mobile Xu Xia,Menghan Yu,Heng Wang,Wen Qi China Telecom Bingming Huang,Jun Liao China Unicom Yang Yang,Liantao Wu,Kai Li,Chenyu Gong,Mulei Ma ShanghaiTech University Chengchao Liang,Lun Tang,Rong Chai,Guozhong Wang Chongqing University of Posts and Telecommunications

11、Gang Shen,Chenhui Ye,Kaibin Zhang,Fangfang Gu Nokia Shanghai Bell Yannan Yuan,Bule Sun vivo Mobile Communication Co.,Ltd.Lanlan Li,Jianjie You Purple Mountain Laboratories Ye Ouyang,Da Wang,Yang Bai,Yan Zhao AsiaInfo Technologies Co.,Ltd.Ling Su,Dandan Hao Ericsson Tingting Yang,Jiahong Ning,Zhengqi

12、 Cui Dalian Maritime University/Peng Cheng Laboratory 4 1 Introduction Native AI is considered to be one of the core architecture features of the future 6G mobile system 1.With the rapid development of AI technologies and applications,their scientific concepts,paradigms/modes,algorithms/models,and m

13、ethods/means need to be more closely and deeply embedded into the architecture,NEs,and functional processes of the new 6G mobile system.In addition,they need to leverage the advantages of the new 6G mobile system as a powerful and ubiquitous telecom infrastructure platform so as to further manifest

14、the performance/effectiveness gains and benefits of AI,and comprehensively help realize the visions of the intelligent connection of everything,intelligent association of services and tasks,and inclusive intelligence in the 6G era.This white paper briefly introduces the background and motivations of

15、 native AI,and systematically elaborates numerous technical requirements(in terms of function,performance,service,architecture,etc.)involved in the design of native AI for the new 6G mobile system,covering three basic AI elements:AI computing power,AI algorithm,and AI data.Further,this paper further

16、 analyzes and summarizes the comprehensive impacts and requirements of native AI on the architecture and standardization of the new 6G mobile system based on the latest research.5 2 Applications Integrating AI and 5G Mobile Systems AI,especially machine learning(ML)and deep learning(DL),was first in

17、tegrated with the 5G system(5GS)according to 3GPP specifications.However,AI service applications and related capabilities and services,for example,how to optimize a specific communication working mechanism based on new AI paradigms,were not fully considered during the design(referred to as the nativ

18、e phase)of the 5GS.The 5GS mobile system has shown tendency toward cloud native,software-defined networking(SDN),and service virtualization.For example,it supports Service Based Architecture(SBA)CN and cloud-based CN.In addition,alliance organizations such as O-RAN and OpenRAN have been actively pro

19、moting the openness,cloudification,and virtualization of wireless networks.Nevertheless,the Next Generation Radio Access Network(NG-RAN)subsystem of the 5GS retains the traditional siloed CT-based base station architecture and relatively rigid RAN protocol stack models due to restrictions in terms o

20、f technology maturity,security,and system O&M complexity.With the gradual penetration and proven benefits of AI functions and services,the 5GS can add the network data analytics function(NWDAF)a new logical function node and various module-level external plug-in/add-on AI functions to the CN based o

21、n the existing system architecture and protocol stack to further enhance system performance and external service capabilities.Each external plug-in/add-on AI functional module is developed for specific known communication issue,such as slice quality assurance,quality of experience(QoE)optimization,m

22、obility prediction,fault locating,and network planning,optimization,and O&M.These AI functional modules are added to improve the performance of the 5GS,optimize wireless transmission efficiency,and simplify network management and O&M.2.1 Status Quo of AI for 5G Many practices have proven that AI,ML,

23、and DL are feasible and effective methods for solving multi-dimensional complex and computing-intensive problems in traditional wireless communication.Currently,AI,ML,and DL have been preliminarily applied and verified at multiple levels and in multiple business domains of the 5GS.Prominent examples

24、 include:1)An intelligent platform helps simplify 5GS O&M,optimize functions such as service scenario identification,network anomaly detection,and fault root cause analysis,and increase energy saving.2)An intelligent module optimizes the deployment of network policies and resources and improves the

25、accuracy of parameter settings.3)Intelligent channel state information(CSI)compression is used to provide feedback on resource overheads.4)Intelligent modulation and coding help improve air interface resource utilization.5)User experience is improved by intelligently identifying and predicting netwo

26、rk traffic distribution,user trajectory,and user behavior 2.All the preceding examples can be called AI for Network(AI4NET).They demonstrate the value of AI as an advanced technology to empower and optimize the 5GS in specific aspects.AI algorithms/models in the 5GS are applied to solve specific kno

27、wn communication issues.That is,these models have undergone a lot of dedicated offline training and verification.Although they only help mitigate or solve specific issues,they also improve communication service performance and reduce system O&M costs to a certain extent.In general,the current AI4NET

28、 still needs improvement in terms of systematicness,thoroughness,globality,and explainability.The AI capability extensibility,iteration enhancement,and AI model generalizability are greatly restricted.In addition,various AI resources and capabilities(covering AI computing power,algorithms,and data)i

29、n the current 5GS are not open or service-oriented.Most of them are used only within the system.2.2 Status Quo of 5G for AI Currently,various mobile AI applications(e.g.,AI recognition,classification,and processing of voice,images,videos,and data)are provided for end users in localized or centralize

30、d cloud AI service mode(e.g.,the public clouds represented by Amazon AWS,Microsoft Azure,Google Cloud,and Alibaba Cloud).Therefore,the 5GS acts more as an underlying data transmission pipe.The 5GS transmits data flows related to AI model training and AI service applications as common user service da

31、ta.The large amounts of sample data required for cloud AI model training are transmitted in the form of application-layer data in the 5GS in an E2E manner.In addition,3GPP Release 16 has defined the NWDAF and related interaction 6 interfaces to implement intelligent applications inside the network a

32、nd empower AI internally and externally,thereby reducing the dependency on traditional cloud AI.All the preceding examples can be called Network for AI(NET4AI),which reflects the benefits of the 5GS to AI businesses and services.For NET4AI intelligent applications,the requirements on KPIs related to

33、 network connection performance in specific scenarios have been specified from 3GPP TS 22.261 3.However,in the preceding NET4AI mobile application examples,all operations in upper-layer cloud AI mode are almost transparent to the 5GS,and cross-layer deep integration or cooperation between upper-laye

34、r AI applications and lower-layer network pipes is not realized.In addition,the communication,sensing,and computing resources,capabilities,and data in the 5GS are not fully open to and utilized by intelligent NEs such as the upper-layer cloud AI application servers or NWDAFs,and are not well adapted

35、 to upper-layer AI service applications.The computing power,algorithm,and data resources owned by each logical NE node in the 5GS are not fully scheduled or utilized by external AI function entities.Most of these resources,such as large amounts of baseband computing power and wireless data in base s

36、tations,serve only traditional communication if not left idle.It is expected that the new 6G mobile system will natively support the capabilities of fully utilizing the computing power,algorithm,and data resources and capabilities of new 6G networks,maximizing the value of these resources to busines

37、ses and services,and further expanding the profitability system of mobile operators.2.3 Status Quo of 5G Network AI Although AI was not natively considered during 5G network design,5G Network AI,that is,integrating AI and 5G networks for mutual benefits,has been continuously explored and continues t

38、o evolve.Standards organizations and forums such as 3GPP,ITU-T,ETSI,and TMF have progressed in the exploration of 5G Network AI.3GPP has made preliminary explorations into 5G Network AI.The status and progress are as follows:In 3GPP Release 16,a new logical function entity,NWDAF,was added to the net

39、work architecture of the 5GS,as shown in Figure 2-1.For details,see 3GPP TS 23.288 4.NWDAFs interact with other function entities in the 5GC(e.g.,AF,PCF,AMF,and SMF)to provide multiple types of network data analysis services.The offered services include receiving network data analysis requests from

40、each network function(NF)entity in the CN,collecting the requested network data,leveraging AI algorithms for data analysis and inference,and returning the network analysis result information to the requesting NF entity.Each NF monitors the operating statuses of the 5GS network and devices based on t

41、he network analysis result information provided by NWDAFs,and performs closed-loop control and optimization on communication services.As of 3GPP Release 17,NWDAFs can analyze network service experience,network performance,slice load,NF load,UE mobility/communication/anomaly events,quality of service

42、(QoS)sustainability,and user data congestion.Figure 2-1 NWDAF-based 5G network data analysis architecture(See 3GPP TS 23.288.)The architecture and functions of 5GS network data analysis were enhanced in 3GPP Release 17,including:logical function splitting and interaction within an NWDAF,collaborativ

43、e data training and model sharing among multiple NWDAF instances,and addition of new function entities to improve data collection efficiency and enhance real-time performance.The NWDAF specified in 3GPP Release 17 has the following features:7 NWDAF entity deployment is more flexible.The centralized,

44、distributed,and hybrid(centralized+distributed)deployment modes are supported.Multiple NWDAF entities can collaborate with each other(e.g.,during analysis aggregation,analysis transfer,and AI data/model sharing).The NWDAF can be further decomposed into model training logic function(MTLF)and analytic

45、 logic function(AnLF).An MTLF can provide ML models for NWDAF entities other than the NWDAF to which the MTLF belongs.The data collection control function(DCCF),analytic data repository function(ADRF),and related data collection optimization processes are added.NWDAF entities can collect data from U

46、Es.Edge computing service experience and network performance can be analyzed.UEs and sessions can be analyzed from the following aspects:slice loads,discrete distribution of data/signaling,WLAN performance,user plane performance,session congestion control,and redundant transmission.3GPP Release 17 a

47、lso researched on and standardized the intelligence of the 5GS management plane.The management data analytics(MDA)function was added,as shown in Figure 2-2.For details,see 3GPP TS 28.104 5.MDA collects data related to network and service events and status.Such data includes network performance measu

48、rement data,trace/Minimization of Drive Tests(MDT)/QoE reports,alarms,configuration data,network analysis data,and AF service experience data.MDA performs data analysis based on specific AI algorithms,generates analysis result reports,and performs network management operations based on the analysis

49、result reports,thereby implementing automatic and intelligent network management and O&M.Figure 2-2 MDA function and service framework(See 3GPP TS 28.104.)3GPP Release 17 is also researching on RAN intelligence.For details,see 3GPP TR 37.817 6.The potential application scenarios include network ener

50、gy saving,load balancing,and UE mobility management.8 3GPP Release 18 continues the research on the performance requirements of AI/ML model data transmission in the 5GS.For details,see 3GPP TS 22.261 3.The 5GS provides QoS guarantee for AI/ML model data transmission based on AI/ML service or applica

51、tion requirements.The 5GS can also monitor the transmission status(e.g.,transmission rate,latency,and reliability)of AI/ML model data and report the status to AI application servers so that the servers adjust AI application-layer parameters accordingly.Currently,3GPP Release 18 is still in the initi

52、ation phase in 3GPP SA2.It considers network intelligence from the following two aspects:AI4NET:focuses on 5GC NE-related analysis and research on potential architecture enhancements,new scenarios,etc.The research on AI4NET includes whether and how to enhance the 5GC architecture to support federate

53、d learning and online learning,report UPF service data to the NWDAF for intelligent analysis,utilize the analysis suggestions from the NWDAF,enhance data collection and storage,support NWDAF-assisted UE route selection policy(URSP),etc.NET4AI:focuses on the performance requirements of AI/ML model da

54、ta transmission in 3GPP Release 18 in SA1,explores 5GS-assisted AI/ML service transmission,supports AI/ML model distribution,transfer,and training,and applies to different AI applications,such as video/speech recognition,robot control,and automobiles.The research also covers the possible architectur

55、e and function extensions that support application-layer AI/ML,possible QoS and policy enhancements,and how the 5GS assists federated learning between UE clients and ASs.ITU-T SG13 has also explored the integration of 5G networks and AI.In November 2017,it established the Focus Group on Machine Lear

56、ning for Future Networks including 5G(FG ML5G).In July 2020,the FG ML5G concluded the work of the second phase and submitted 10 technical specifications to SG13,including specifications on AI use cases,architectural framework,intelligence levels,data processing,ML function orchestrator,service frame

57、work,etc.In addition,the FG ML5G proposed an ML-oriented management subsystem,and the concept of multi-domain,multi-cloud,and multi-level ML workflows based on the functions required in different phases of the ML lifecycle.ETSI has also explored the integration of ICT systems and AI.As early as Febr

58、uary 2017,ETSI established the Experiential Networked Intelligence(ENI)Industry Specification Group(ISG).The ENI ISG defines an AI engine that provides intelligent services for applications such as network O&M,service orchestration,and network assurance.Figure 2-3 shows the functional architecture o

59、f ENI 7.9 Figure 2-3 Functional architecture of ENI Currently,the ENI system contains knowledge management,model management,policy management,and other modules.After data is processed by AI modules,the ENI system can automatically provide service operation and assurance,slice management,and resource

60、 orchestration for networks.ENI functions are continuously evolving.For example,the ENI system now supports intent-driven networks.TMF focuses on OSSs/BSSs when exploring the integration of 5G networks and AI.Currently,TMF is carrying out an AI and Data Analytics(AI&DA)project,which focuses on the a

61、rchitecture,use cases,AI terminology,data processing,and AI training.Table 2-1 describes the research directions and specific work of this project 8.Table 2-1 Major work involved in the AI&DA project of TMF 10 2.4 Challenges in Integrating AI and 5G Mobile Systems The 5GS network is expected to bett

62、er utilize AI capabilities to enhance the network itself and support AI service applications,first in the CN and then in the RAN.The NWDAF introduced in 3GPP Release 16 aims to improve AI data collection and analysis capabilities.For example,it can provide analysis result information for other core

63、NFs and UEs,helping optimize network service provisioning.The NWDAF also supports data collection from the network O&M and management system in the 5GS,and provides dedicated services for registering NFs and opening metadata.Even so,the integration of the 5GS and AI faces the following challenges:Li

64、mited data sources:NWDAF entities mainly collect and analyze the data received by 5GC NFs.However,in a broad sense,the data from the wireless infrastructure,environment,devices,and various sensors is not fully considered.As a result,AI data samples are incomplete.Transmission bandwidth consumption:A

65、 large number of transmission bandwidth resources are consumed to collect AI data regardless of centralized or distributed NWDAF deployment.If data sources are far from NWDAF entities,there may be latency in data updates.Lack of data privacy protection:The NWDAF collects and analyzes data in a centr

66、alized manner.Data sources usually come from the same service domain.Therefore,data privacy protection is not fully considered in architecture design,and user privacy may be compromised.Lack of support for external AI services:As an internal function of the 5GC,the NWDAF is mainly used to enhance an

67、d optimize the 5GS.External AI applications cannot be directly served by or benefit from the AI functions of the 5GC or RAN.Insufficient infrastructure utilization:Key 5G functions,such as network slicing,ultra-reliable low-latency communication(URLLC),and massive machine-type communications(mMTC),a

68、re designed to meet vertical industry requirements on performance,function,and operation.However,the support for native AI(data governance and services,distributed architecture,etc.)is not specifically considered during the architecture design of these key 5G functions.Various resources(that have lo

69、w utilization or are in an idle state)in the network infrastructure are not fully utilized by service applications and the value of these resources is not effectively realized.Lack of data governance and services:AI involves more than data collection and training&analysis and inference.To support 6G

70、 native AI,the architecture of AI data governance and services needs to be systematically designed.This approach was not considered during the design of the 5GS.The value of data services needs to be further manifested in the future.11 3 Mechanism of Native AI in the New 6G Mobile System The operati

71、on data information communication technology(ODICT)industry believes that native AI will become one of the core features of the new 6G mobile system in the future 1.Therefore,deep integration with AI is comprehensively considered in the requirement and architecture design phases of the new 6G mobile

72、 system.Different from adding AI functions to the 5GS by means of add-ons,patches,and plug-ins,the goal of native AI will bring many profound impacts and challenges to the design of the new 6G mobile system.This section will first elaborate the motivations and reasons for implementing 6G native AI(e

73、.g.,why native AI can better adapt to new 6G scenarios and use cases in the future,and how to form new businesses and generate new value).3.1 Necessity Analysis With the evolution toward future mobile networks,network management and O&M need to transform from local intelligent operations for lower c

74、ost and higher efficiency,to E2E high-level network autonomy.However,the R&D of existing AI use cases is generally performed by means of patches,external plug-ins,and siloed add-ons,and lacks a unified system framework.Most AI models lack the means to pre-verify the effect of applying the models and

75、 QoS assurance.AI model training and learning are decoupled from analysis and inference in that they are not performed at the same time.The effect of an AI model can be verified only after AI inference is finished,and this can have a great impact on the live network.As a result,high-level network au

76、tonomy is impossible,and AI models cannot be pre-verified,evaluated online,or fast optimized in an automatic closed-loop manner.In addition,AI model training/retraining requires a large amount of sample data,but it is difficult to collect data in a centralized manner.This leads to a number of issues

77、,for example,high network transmission overhead and training overhead,long iteration and update time,slow convergence,and poor generalizability of AI models.Therefore,the new 6G mobile system needs to further enhance the training and application performance of AI models so as to improve the network

78、autonomy level.The key driving forces of native AI include assisting various vertical industries in intelligent digital transformation,exploring new intelligent business models,and providing new 6G scenarios and capabilities.On the premise of protecting data privacy of vertical industries and preven

79、ting data from being transferred out of campuses,the new 6G mobile system must be able to provide distributed and regional computing resources,platforms,and services to flexibly provision intelligence capabilities on demand anytime,anywhere,and enable data-centric computing.Compared with traditional

80、 cloud AI service providers,6G native AI can provide intelligence capabilities and services with higher timeliness,privacy,and performance.In addition,6G native AI can provide inter-industry federated intelligence to facilitate cross-domain and cross-industry intelligence convergence and intelligent

81、 digital sharing.With the evolution of intelligent devices in the future,massive devices will generate more data,and the computing and intelligence capabilities of devices will become increasingly stronger.6G native AI needs to coordinate network AI and device AI to provide ToC users with the ultima

82、te service experience and higher-value novel data information communication technology(DICT)services.Ensuring the security and trustworthiness of future networks is also an important research topic.Native AI can promote the realization of native network security and trustworthiness,and autonomously

83、detect and defend against various types of potential attacks and threats.3.2 Feasibility Analysis The successful integration of the 5GS and AI has proven the feasibility of implementing network intelligence to some extent.This section analyzes the feasibility of deeply integrating native AI and the

84、new 6G mobile system from the perspective of three key AI elements.Computing power:Due to the requirements on latency,reliability,data security,and privacy protection,deploying computing capabilities and resources on the network edges has become a major trend in the 5G era.In the 6G era,data connect

85、ion and computing may be further integrated.12 For example,a dual-infrastructure that integrates both connection and computing may emerge,which can provide a computing service-related basis for integrated native AI design.Algorithm:Although centralized data processing on the AI cloud has its advanta

86、ges,issues related to data privacy,high performance,and computing energy consumption remain challenging.If the new 6G mobile system can integrate algorithms/models and intelligence capabilities into the network,that is,if data and tasks can be processed intelligently no matter where they are,then an

87、 effective alternative to centralized data processing on the AI cloud is possible.As data,computing resources,and computing capabilities are moved to network edges,AI algorithms/models will also be moved to network edges(e.g.,edge nodes and base stations)for execution and maintenance.Data:Traditiona

88、l communication networks mainly function as data transmission pipes.They do not actively generate or process service data,except network management and operation data.Such data limitations may be one of the reasons why 5G AI focuses on network performance improvement,optimization,and automatic manag

89、ement and O&M.In the 6G era,as sensing technologies,industry digitalization,and edge computing become mature,new 6G networks will themselves become huge wireless sensor networks capable of actively generating and processing massive heterogeneous data.This provides integrated 6G native AI design with

90、 a foundation related to data services,for example,the collection and preprocessing of massive training data samples,and on-demand data asset transfer.3.3 Gain Analysis 6G native AI will greatly strengthen the coupling between AI resources(including the computing power,algorithms,and data)and the ne

91、w 6G mobile system.The so-called communication,computing,and intelligence approach will evolve to the communication+computing+intelligence convergence approach.In this way,the integration and reuse rates,utilization efficiency,and comprehensive performance of various resources in the new 6G mobile s

92、ystem will be improved,offering more cost-effective service applications.End users can enjoy more ubiquitous and cost-efficient computing intelligence services.Compared with the traditional working mode of centralized cloud AI server+intelligent edge nodes,the native AI mode enables AI resources to

93、be more widely,evenly,and flexibly distributed on the ubiquitous infrastructure platform of the new 6G mobile system.AI computing intelligence operations are closer to data sources,task sources,and end users,and can more efficiently adapt to dynamic conditions of air interfaces(e.g.,user environment

94、 and channel changes,and network topology and resource updates).Therefore,it is easier for the new 6G mobile system(especially base stations in the system)to perform fast,precise closed-loop optimization as well as real-time policy adjustment and trend prediction in the user-oriented dynamic environ

95、ment.In distributed AI machine training mode represented by federated learning,the decentralized native AI mode is more conducive to the privacy protection of user data,balancing among data and computing tasks,and enhancement of local security and autonomy of networks/subnetworks.In terms of collect

96、ion,processing,transmission,and transfer of AI data(e.g.,algorithms/models,training samples,basic parameters,and feature parameters),native AI may rely on dedicated logical functions such as the data plane and intelligence plane of the new 6G mobile system to perform more efficient,flexible,and robu

97、st data transfer and sharing.In this way,the AI data transmission latency,transmission resource consumption,and system energy consumption are all reduced.The further standardization of 6G native AI technologies can also promote the interconnection and collaboration among ODICT vendors in communicati

98、on+computing+intelligence devices,functional modules,and AI task processes,and even reconstruct business forms and models in the future.This will also further encourage more industry participants to jointly build a broader,more secure and trusted AI resource capability service platform,thereby reali

99、zing the 6G visions of ubiquitous AI and inclusive intelligence.13 3.4 Analysis of New Features and New Paradigms The new 6G mobile system will face immense technical challenges to implement native AI.The details are as follows:Highly differentiated quality of artificial intelligence service(QoAIS)r

100、equirements:On the one hand,the operating status of wireless networks fluctuates greatly,and user and service requirements are highly dynamic,especially the requirements of vertical industry users.On the other hand,existing mobile systems lack systematic evaluation and assurance for QoAIS.Limited co

101、mmunication and computing resources:The computing power and intelligence of devices on network edge nodes are insufficient.The storage capability is poor,and the transmission bandwidth between radio access nodes and edge devices is limited.In addition,heterogeneous resources(e.g.,computing power,dat

102、a,and connection)related to AI services cannot be systematically managed,controlled,or allocated.Lack of simulative AI training and verification environment:On the one hand,simulation environments have high requirements on the real-time performance of data synchronization and the data volume.On the

103、other hand,in simulation environments,pre-verification of the effect of applying the AI models and QoS assurance are required to prevent potential negative impacts to live networks after the AI functions are enabled.Requirements of vertical industries:Native AI needs to provision and flexibly orches

104、trate AI resources,functions,and services to various device-edge-network-cloud nodes,and provide a distributed AI architecture that enables close device-network collaboration and a QoAIS evaluation and assurance system.In addition,native AI needs to enable inter-industry cross-domain federated learn

105、ing and a framework for sharing knowledge,experience,and data to facilitate integration with the digital twin networks mapped to the industry service logic.Future-proof network autonomy:Native AI needs to implement the following functions required by network autonomy:self-discovery,self-orchestratio

106、n,self-configuration,self-optimization,effect self-evaluation,multi-domain closed-loop management,AI effect pre-verification,online evaluation,and closed-loop fast optimization.Integration of heterogeneous communication,storage,and computing resources enables efficient utilization of various resourc

107、es on the network.Secure and trusted network autonomy solutions need to be supported in the future.Common end users:Native AI needs to intelligently sense device data and computing resources,leverage the massive data of a massive number of devices and computing resources to implement a distributed A

108、I architecture that supports device-network integration,and perform anonymization,association,and aggregation on user data in order to guarantee the security of user data and privacy.Given the preceding technical challenges,6G native AI must not only achieve high-level network autonomy,but also prov

109、ide high-quality and guaranteed AI services for numerous vertical industry users and end users.The new 6G mobile system must have at least the following features of new technology paradigms:E2E lifecycle orchestration and management of QoAIS:An AI/ML service quality evaluation and assurance system n

110、eeds to be built to orchestrate resources such as AI/ML algorithms/models,computing power,and data throughout the lifecycle.Deep integration of computing and communication based on native AI:In scenarios where data,computing power,bandwidth resources,and transmission latency are restricted on the ne

111、twork,native AI needs to consider joint orchestration of computing and communication resources.The 6G network architecture,protocol orchestration,and function processes will be reconstructed to fully adapt to 6G air interfaces and network-side transmission features and optimize the performance of na

112、tive AI/ML models.Integration of native AI and digital twins:Native AI can provide AI/ML models required by digital twin networks,implement data augmentation and self-generation,and lower the requirements of digital twin networks for physical network data collection.Digital twin networks can pre-ver

113、ify and 14 optimize the effect of AI/ML workflows or AI models to prevent negative impacts to the live network after AI/ML functions are added.15 4 Technical Requirements of Native AI in the New 6G Mobile System 4.1 Capability Requirements The new 6G mobile system will be a super wireless infrastruc

114、ture platform that deeply integrates communication,sensing,computing,intelligence,and storage.Internally,the system will offer more powerful capabilities in communication,sensing,computing,intelligence,and storage,for example,more functions and better performance(indicated by KPIs).Externally,it can

115、 provide comprehensive businesses,services,and applications in terms of communication,sensing,computing,intelligence,and storage to consumers,businesses,and third-party customers.Therefore,the new 6G mobile system is not only a super wireless data transmission pipe,but also a huge distributed radar

116、sensor network,ubiquitous heterogeneous computing platform,and AI server array.6G native AI is closely related to the features of the new 6G mobile system.The following will describe the basic capability requirements of 6G native AI from three aspects:computing power,algorithm,and data.4.1.1 Computi

117、ng Power Requirements As a new form of productivity for the 6G native AI platform,computing power is a solid foundation for supporting all data capabilities and services.Currently,the rapid development of 5G,big data,and AI technologies are continuously driving explosive data growth and increasing t

118、he complexity of AI algorithms.As a result,higher computing scales and capabilities are required.For the next-generation 6G computing and network integration system that integrates computing,communication,sensing,and AI capabilities,computing resource discovery,sensing,measurement,on-demand scheduli

119、ng,and openness will become the major trends of future mobile network development.Computing resource discovery:The new 6G mobile system involves collaboration among technologies such as cloud computing,fog computing,and edge computing.These technologies vary in terms of capabilities and coverage sco

120、pe,and can be separately applied to IoT applications and services at the region,locality,and device levels.New 6G networks must be able to discover and register multi-level heterogeneous computing resources in time to support basic functions such as real-time data sensing,processing,control,and exec

121、ution.As a basic function of the new 6G mobile system,the computing resource discovery and registration capability serves as the foundation of computing services.Computing requirement sensing:As more and more data is generated and computing algorithms become increasingly complex,the computing requir

122、ement sensing capability of the new 6G mobile system needs to become more intelligent.Typical computing service requirements shift from simple data sensing,collection,and representation to extraction and analysis of computing service requirement information.The new 6G mobile system can be widely use

123、d in scenarios such as environment monitoring,city management,and health care.The tactile network enabled by the new 6G mobile system will facilitate real-time sensing of computing requirements and enable efficient data processing,information extraction,analysis,and decision-making.Unified computing

124、 power measurement:The computing capabilities involved in the new 6G mobile system are classified into the following three categories 9:Logical computing capability:A general-purpose basic computing capability.A CPU is an example of a microchip with such computing capability.Parallel computing capab

125、ility:A highly efficient computing capability dedicated to accelerating the processing of data types such as graphics and images.A GPU is an example of a microchip with such computing capability.Neural network computing capability:A computing capability used to accelerate ML and neural networks.An N

126、PU is an example of a microchip with such computing capability.16 For a variety of devices and platforms that provide heterogeneous computing(computing requirements of multi-vendor chips,diversified computing types,and different users),quantifying heterogeneous computing resources in a unified manne

127、r is the foundation for scheduling and utilizing computing resources.To this end,the new 6G mobile system needs to provide measurement functions or methods that map heterogeneous computing resources to a unified dimension.On-demand computing resource scheduling:The 6G native computing and network in

128、tegration system will combine the advantages of edge computing,fog computing,network cloudification,and intelligent control,and implement extensive computing resource management and dynamic on-demand scheduling through powerful network connections.Different from the traditional centralized managemen

129、t or intensive provisioning of cloud computing resources,the resource management module of the 6G native computing and network integration system will consider the impacts of network status(e.g.,air interface information,network latency,and network loss)on distributed computing resource scheduling t

130、o achieve the optimal performance and resource utilization.Therefore,the core capability of new 6G networks will be to improve the efficiency of ubiquitous distributed computing.Computing and network integration aims to improve the efficiency of intelligent computing,communication,and AI services.Co

131、mputing power openness and transaction:The native computing power of the new 6G computing and network integration system will no longer be provided only by mobile operators or cloud service providers.Any nodes that can contribute idle computing power,such as base stations,mobile phones,computers,gam

132、e consoles,and available small data centers of enterprises,can become new types of computing resources.In consideration of factors such as security,cost,efficiency,and reliability,the new 6G mobile system needs to define reasonable mechanisms(e.g.,blockchain technology)to manage ubiquitous computing

133、 resources and the open the transactions of such resources 10.4.1.2 Algorithm Requirements As one of the three AI elements,algorithms(models)are one of the decisive factors for new 6G networks to implement native AI and offer various AI services.The algorithm requirements for 6G native AI cover at l

134、east the following aspects:AI algorithm indicator requirements:Currently,there are various types of AI algorithms and models,which are iterated quickly.The AI algorithms/models differ greatly in terms of functions,performance,execution time,and computation complexity.From the perspective of native A

135、I and external QoAIS of new 6G networks,the indicator requirements and indicator descriptions of AI algorithms for 6G need to be defined in a standardized manner,and the quantification scope of each indicator needs to be provided for each AI service.For example,the algorithm requirements can be comp

136、rehensively described from the perspectives of benefits,computation complexity,network types,execution time,input and output requirements,data dependency,and generalizability of AI algorithms/models.AI algorithm indicators that cannot be easily quantified can be described hierarchically.AI model tra

137、ining requirements:Generally,AI models can converge only after long-term multi-round training based on large amounts of data.In a 6G native AI system,the training mode of a specific AI model needs to be decided in real time based on objective requirements(e.g.,volume of collected system data,computi

138、ng resources,and privacy assurance)and resource status(e.g.,current status of communication and computing resources in the system).Potential training modes include offline training,online training,federated training,etc.AI model training must optimally match the real-time resource status of the new

139、6G mobile system.AI model description and interaction requirements:The new 6G mobile system contains diversified types of AI models to serve different AI use cases and services.These models should be able to interact with and be shareable among NE nodes in different network architectures.Therefore,a

140、 description language(or method)of AI algorithms/models needs to be defined for 6G,and a unified process needs to be specified to ensure that all AI algorithms/models are interactive and executable.17 AI algorithm evolution requirements:Wireless communication and AI algorithms are two fast-growing a

141、reas.The architecture and capabilities of AI algorithms/models need to be evolvable.AI algorithm evolution can be evolution at the parameter level(performance self-optimization)or evolution at the model architecture/structure level(architecture self-growth).4.1.3 Data Requirements The native AI on n

142、ew 6G networks is mainly data-driven.Therefore,from the perspective of AI data requirements,it is crucial for 6G networks to have at least the following key capabilities:Data collection capability:Future networks will have abundant data,including infrastructure resource information,customer and part

143、ner information in BSSs,industry information in industry communication systems,and information about end users and environments.The network data collection function is required to collect such data efficiently and securely.The process of collecting network data includes establishing secure connectio

144、ns with various data sources,determining the collection scope and method,storing the collected(and preprocessed)data in the database,and performing O&M on the database.Data analysis capability:Data mining and ML methods will need to be used to analyze and extract large amounts of collected data in o

145、rder to make maximum use of data functions,extract key information and knowledge from data,and provide customers with desired data services.For example,the statistical features of historical data can be leveraged to automatically detect network faults,service impairments,and network anomalies.User d

146、ata can be utilized to predict future events,and analyze user behavior and preferences to better serve users.Diverse data analysis and mining algorithms are available,and algorithms characteristics vary depending on data types.The native AI of new 6G networks needs to be capable of selecting the mos

147、t appropriate data analysis and mining algorithms for each type of service data.Data privacy and security protection capability:Collecting and storing sensitive data involves privacy disclosure risks and therefore requires a high level of privacy protection.Data anonymization is an important action

148、to address privacy concerns and achieve legal compliance,and this is particularly important for enabling secure AI data services in the new 6G mobile system.Data anonymization in AI model training and inference has been attracting a lot of attention.In recent years,a lot of research has been conduct

149、ed in AI technology fields such as differential privacy,homomorphic encryption,and multi-party security.Data storage capability:In the future,data storage will experience a core-to-edge transformation,and data will be stored on cloud,edge,and device nodes in a distributed manner.In addition,unstruct

150、ured,semi-structured,and structured data storage architectures will need to be supported.The new 6G mobile system must be able to store data by layer and by domain to provide multi-layer fault tolerance and redundancy capabilities.For example,high-value detailed data and essential lightweight summar

151、y data can be stored on the cloud,whereas low-value detailed data and other lightweight summary data can be stored at the edge.Take the edge scenario as an example.The detailed data stored locally can be quickly processed to empower local applications.The resulting data is then transmitted to the cl

152、oud or other edge nodes based on storage and usage specifications.Data exposure capability:Aggregated internal and external data will be encapsulated into external data services after integration,privacy protection,and standardization processing.Such external data services are open to 6G native AI n

153、etworks and external third-party users in the form of standardized data services(including data sets,AI model data,and prediction services).Based on 6G application scenarios,data services can be classified into Internet of Humans(IoH)data services(e.g.,location,online behavior,and call drop rate of

154、users),IoT data services(e.g.,device,equipment,and vehicle),and other data services.4.1.4 Other Requirements In addition to meeting the preceding requirements for AI computing power,AI algorithms,and AI data,6G native AI needs to further promote the implementation of intrinsic security(proactive imm

155、unization)in the new 6G mobile system.For example,in the face of more unknown and uncertain service scenarios and 18 user environments in the future,the new 6G mobile system must be able to support the following functions based on native AI capabilities:real-time sensing,analysis,inference,and predi

156、ction of various potential threats and risks,comprehensive autonomous immunity,active defense,and multi-node coordinated security defense policies.Various new variants of security threats and risks can be identified as early as possible based on self-learning training and inference of native AI on n

157、ew 6G networks.Therefore,security cognition and experience synchronization can be performed in real time across the 6G network,achieving all-round coordinated defense and eliminating risks.Currently,the 5GS has achieved decent energy saving gains with the assistance of plug-in AI functions.For insta

158、nce,for a 5G network in a specific region,20 million kWh of electricity can be saved per 10,000 sites each year,significantly reducing electricity costs while maintaining stability in network KPIs and user experience 11.Nevertheless,the energy consumption of the future new 6G mobile system can be fu

159、rther reduced.In the future,6G native AI needs to continuously improve the energy efficiency and lower the power consumption of the new 6G mobile system to effectively support the new energy and double carbon(carbon peaking and carbon neutrality)strategies.4.2 Service Requirements Compared with the

160、internal AI capabilities of the new 6G mobile system,the external businesses and services provided by the new 6G native AI system are more important and valuable to users and customers.In the future,the multi-subject fields served by the new 6G mobile system include:individual(ToC),family(ToH),enter

161、prise(ToB),industry(ToI),and society(ToS).The subject domains or levels have different requirements for 6G native AI businesses and services,for example,they have different AI capability requirements,as described in section 4.1.The following describes the service requirements from three aspects:comp

162、uting power,algorithm,and data.4.2.1 Computing Service Requirements The rapid growth of AI-intensive computing requirements will further spur the rapid development of computing services of the new 6G mobile system.In the future,the new 6G mobile system will need to provide computing,data,and algorit

163、hm services required by internal and external AI applications,enabling open and shared computing capabilities and AI services.The evolution requirements of computing and network integration must be considered during the design of multi-layer computing and network integration,which will serve as the

164、next-generation network architecture.To implement ubiquitous computing power connection and global optimization of AI computing power on the network,the following functions should be supported:on-demand,flexible scheduling and utilization of computing power,and reasonable distribution of computing s

165、ervices.The goals of multi-layer computing and network integration are as follows:The network can sense and manage ubiquitous computing services.Users do not need to be concerned about the status of computing resources on the network.Computing power is dynamically deployed and flexibly scheduled to

166、services that have computing requirements.To meet the computing service requirements,the new 6G mobile system must be able to provide the following types of computing services:Latency-sensitive services:The intelligent sensing and decision-making functions of the new 6G mobile system must provide ab

167、undant computing power for AI services in real time to accelerate transmission and computing,thereby shortening latency.Compared with the 100 ms latency required by mobile cloud computing(MCC),the new 6G mobile system based on multi-layer computing and network integration can meet the ultra-low late

168、ncy(110 ms)requirement.Application scenarios include immersive virtual reality(VR)and extended reality(XR).Energy consumption-sensitive services:Users or customers can allocate the computing power of the new 6G mobile system to migrate energy-intensive AI computing tasks to new 6G networks for distr

169、ibuted parallel computing.Because computing is not performed locally,there is a significant 19 amount of energy savings.Application scenarios include emergency rescue and smart wearable devices.Privacy-sensitive services:Compared with traditional centralized computing,the computing services of the n

170、ew 6G mobile system enable user data to be transmitted to the cloud data center without passing through the complex CN.The computing power can be scheduled and allocated on the new 6G mobile system for secure,reliable,and high-privacy computing without requiring users to upload private data.Applicat

171、ion scenarios include inclusive finance and data silos.Service experience-sensitive tasks:Multi-layer computing and network integration has the advantage of providing resource services that are close to users.With this advantage,the new 6G mobile system can use the intelligent sensing function to ac

172、curately predict and determine user computing behavior and requirements,so as to schedule computing resources in advance and provide real-time and effective computing services.Application scenarios include immersive VR,XR,and sensory interconnection.4.2.2 Algorithm Service Requirements New 6G networ

173、ks use AI algorithms/models to implement internal and external native AI services.The quality of AI algorithms determines the quality and performance of native AI applications.The external AI algorithm/model services provided by new 6G networks must support at least the following features:Standardiz

174、ed description and storage of AI algorithms/models:There are many types of AI algorithms/models,and their use cases vary greatly depending on scenarios.The new 6G mobile system needs to provide standardized descriptions of algorithms/models(services),and define algorithms/models and related services

175、 using a standardized description language under a unified framework.The descriptions must be explainable.Configuration parameters and configuration files must be provided for each AI algorithm/model,and content modification must be allowed.The description file of an algorithm must cover the definit

176、ion,configuration,usage,performance,and application scenarios of the algorithm.AI algorithm classification,indexing,and service interfaces:AI algorithms/models need to be classified and indexed.Standardized formatted interfaces must be provided for using AI algorithm services.Multiple search methods

177、,such as keyword-based search,need to be supported.The new 6G mobile system must support the deployment of a single AI algorithm and the joint deployment of multiple AI algorithms in stack mode.Algorithms can be customized during AI algorithm/model deployment.However,a modified algorithm can serve a

178、s a native AI algorithm of new 6G networks only after it passes strict authentication.AI algorithm authentication:The AI algorithms/models used to provide external services can be used as native AI algorithms of new 6G networks only after the algorithms/models pass strict authentication.The authenti

179、cation must involve test,verification,and performance confirmation.The authentication result information must be contained in the algorithm description file.Such information includes but is not limited to application scenarios,performance,data security,data privacy,and resource overhead.Only strictl

180、y authenticated AI algorithms/models can be retrieved,deployed,and used on 6G native AI networks.AI algorithm training:Algorithm training,also called algorithm modeling,aims to search for the optimal model parameters based on the provided data sets.Information about AI algorithm training needs to be

181、 provided in the description file.Such information includes but is not limited to the training data,training environment,training resource overhead,and test result information.As one of the network AI services,AI algorithm training is performed based on specific resources allocated by the new 6G mob

182、ile system.Specific training data sets are used,and the training process can be monitored and controlled.AI algorithm reuse:If the new 6G mobile system separately configures a native AI algorithm/model for each AI use case request,this will result in huge costs and overheads.To lower costs and overh

183、eads,the new 6G mobile system must be able to reuse existing or trained AI models on the network to handle multiple use cases that have similar requirements.Model reuse can be performed 20 at different levels as required,for example,by an AI algorithm/model framework,algorithm/model hyperparameter,o

184、r specific algorithm/model parameter.AI algorithm testing:Test data sets are used to test and evaluate the generalizability and other performance of AI models.The test indicators include but are not limited to accuracy and recall.The new 6G mobile system must provide complete test environments and t

185、est results.AI models that fail to pass tests cannot be deployed.As one of the network AI services,AI algorithm testing is conducted based on specific resources allocated by the new 6G mobile system.Specific test data sets are used,and the training process can be monitored and controlled.AI algorith

186、m inference:After an algorithm/model passes authentication,it can be used for new data inference and decision-making.AI algorithm inference services are provided for authenticated users through standardized service interfaces.The results of AI algorithm inference can be used to further optimize algo

187、rithms/models.4.2.3 Data Service Requirements One of the functions of new 6G networks is to create an intelligent,ubiquitous,connected world and provide AI capabilities and services for applications in various fields based on ubiquitous big data.New 6G networks are designed to integrate AI and big d

188、ata applications to form an E2E architecture.To achieve this goal,new 6G networks need to support independent data-plane functions,build an architecture-level data service framework,and establish knowledge graphs that support native AI,meeting the network-wide all-domain native AI requirements on da

189、ta collection,ML,intelligent services,and application enablement,while protecting data security and privacy.In this way,new 6G networks can provide higher data transmission rates and more reliable transmission links,meeting the requirements for new service types and stricter QoS.The essence of AI ap

190、plications is to fully explore and continuously learn the value of big data based on continuously enhanced computing power.With big data as the foundation 12,6G native AI will continuously drive the development of new 6G technologies and services through four basic steps of ubiquitous data collectio

191、n,processing,transmission,and utilization on networks and devices.In the AI data processing process,the four basic steps can be further decomposed into the following actions:data discovery,collection,preprocessing,transmission,transfer,training,processing,analysis,inference,decision-making,mining,au

192、gmentation,transaction,storage,and privacy protection.These actions run through big data processing and application processes,and correspond to AI data functions and services.See Figure 4-1 13.Figure 4-1 Big data processing and application 21 AI technologies have made rapid development with the supp

193、ort of big data technologies.AI has increasingly high requirements on data quality,scale,and personalization,and demands scenario-specific and domain-specific data.Accordingly,the processing capability in each step(e.g.,data collection,labeling,and feature extraction)of the big data process needs to

194、 be enhanced 14.In the process of building AI analysis models using DL or ML,data sets are collected from various data sources,such as files,databases,and sensors.However,the collected raw data needs to be preprocessed and cannot be directly used for analysis.To obtain better results from ML and DL

195、application models,data must also be represented and stored in appropriate formats.Certain ML and DL models require information in specific formats.For example,the random forest algorithm does not support null values.Therefore,to successfully execute the random forest algorithm,the null values in ra

196、w data sets must be handled prior to algorithm execution.Data augmentation,a common technique used in intelligent network planning and optimization,aims to add training data sets and compensate for the lack of sample data.It can diversify the training data sets and enhance the generalizability of th

197、e trained AI models 15.Feature extraction starts with the collected raw data and builds informative and non-redundant eigenvalues to facilitate subsequent ML and generalization steps.If the amount of input data of an AI algorithm is too large to process and is redundant,complex feature data can be c

198、onverted into a group of simplified features.Features are extracted from data sets based on application scenarios.The feature data satisfying a specific scenario is obtained through syntax processing and semantic analysis 16.Based on the preceding native AI data functions,the new 6G mobile system ne

199、eds to provide various external AI data services on demand,including AI data(source samples,data sets,feature sets,model files,evaluation effects,etc.)and data service items corresponding to each AI data function.Proper layout of storage functional modules and data transmission modes in the network

200、architecture will ensure high-reliability and low-latency transmission of AI data on the network,meeting the requirements of processes such as AI algorithm/model training,verification,inference,and decision-making.AI technologies must intelligently support multiple types of data services.4.2.4 Other

201、 Service Requirements The new 6G mobile system must provide a complete QoAIS evaluation and assurance system to accurately judge,measure,and guarantee the quality levels of AI businesses and services provided by the 6G native AI system.QoAIS corresponds to the concepts of QoS and QoE in traditional

202、communication services.QoAIS is an indicator system for systematically evaluating and guaranteeing the quality of AI businesses and services 17.New 6G networks will build native AI capabilities and form a capability service system(i.e.,artificial intelligence as a service,AIaaS)that applies to vario

203、us intelligent application scenarios.Different such scenarios(e.g.,high-level network autonomy,inclusive intelligence of industry users,ultimate user experience,and intrinsic security)have different requirements on the quality of AI businesses and services.Therefore,an indicator system is needed to

204、express user-level requirements,and orchestrate and manage the comprehensive effect of AI elements(including algorithms,computing power,and data)on the network in a quantitative or hierarchical manner.Native AI services of new 6G networks can be classified into the following types:AI data,AI trainin

205、g,AI inference,and AI verification.Each type of AI services requires an independent set of QoAIS.The QoS concept used for traditional communication services mainly focuses on connection-related performance indicators,for example,transmission latency and throughput(e.g.,MBR and GBR).In contrast,QoAIS

206、 needs to cover aspects such as AI performance,overhead,security,privacy,and autonomy,and comprehensively evaluate the service quality of network native AI from dimensions such as connection,computing power,algorithm,and data.Therefore,compared with QoS,the QoAIS indicator system must first be exten

207、ded in terms of content.Table 4-1 describes some indicator requirements of AI training services.22 Table 4-1 QoAIS indicator system for AI training services AI Service Type Evaluation Dimension QoAIS Indicator AI training Performance Performance bounds,training time,generalizability,reusability,robu

208、stness,explainability,consistency between the loss function and the optimization objective,and fairness Overhead*Storage overhead,computing overhead,transmission overhead,and energy consumption Security*Storage security,computing security,and transmission security Privacy*Data privacy level and algo

209、rithm privacy level Autonomy Full autonomy,partially controllable autonomy,completely controllable autonomy *indicates common evaluation indicators of different types of AI services.QoAIS is an important input of native AI orchestration and management systems and control functions of new 6G networks

210、.Native AI orchestration and management systems need to decompose the top-level QoAIS indicators,convert the decomposed indicators,and then map them to specific QoS requirements in terms of data,algorithms,computing,and connections.To maximize the utilization of knowledge,experience,and data in vari

211、ous industries,the new 6G mobile system also needs to provide a federated learning mechanism both within a single industry and across industries and an external service system for sharing data,knowledge,and experience across industries.4.3 Architecture Requirements To meet the preceding capability a

212、nd service requirements,6G native AI needs to have a mobile system architecture and function system that differ from the traditional communication service-centric ones.For example,the 6G native AI architecture will emphasize the self-generation,self-optimization,and strong utilization of system func

213、tions and data,implement data-centric computing,and highlight distributed,collaborative,and task-/project-oriented network operations.Any intelligent node can be an initiator or terminator of intelligent businesses or services.The following describes the architecture requirements from three aspects:

214、computing power,algorithm,and data.4.3.1 Computing Architecture AI technologies have progressed by leaps and bounds over recent years,and applications are becoming increasingly intelligent and complex.Novel application scenarios proposed by 6G,such as immersive cloud XR,sensory interconnection,and i

215、ntelligent interaction,pose numerous challenges to the O&M and management of traditional mobile communication networks.For example,an extremely large number of devices are connected to the network,and the network scale continuously increases.In the past,computing and communication were independent.T

216、raditional mobile networks lack effective management and integration of computing resources and communication resources.In the future,external security threats will become increasingly severe,and traditional mobile networks are not equipped to provide fine-grained security protection.In addition,6G

217、services will need to be deployed 23 more quickly and agilely,but service deployment and iterative upgrade on traditional networks are time-and effort-consuming.During the deployment of a new 6G network,both traditional O&M issues and the objective of maximizing the utilization of edge computing res

218、ources need to be considered.New 6G networks are expected to provide optimal resource allocation and network connection solutions and achieve optimal utilization of network-wide resources based on network information(e.g.,path and latency)and user requirements by distributing computing power informa

219、tion,storage information,and algorithm information of each service node.Multi-agent sensing,cloud-native computing,and digital twin technologies are also combined to improve the in-communication computing sensing capability of networks.To provide functions such as system management capability extens

220、ion,in-communication computing scheduling,and fine-grained security protection,and deliver the ultimate user experience,new 6G networks will need an innovative network architecture.Figure 4-2 Deployment architecture of the multi-layer computing and network integration system The new 6G network will

221、be a cloud-network-edge-device system supporting multi-layer computing and network integration 18,as shown in Figure 4-2.It will adopt an innovative architecture within which computing,communication,control,and storage are better integrated.A computing and network integration system is a multi-layer

222、 network in which computing and communication nodes can provide native live-network computing power.Integration of computing and communication enables better utilization of computing and communication capabilities.Computing and network integration supports various applications and components.Computi

223、ng resources are properly allocated using the computing power scheduling function.In addition,network computing services and communication services are provided and managed.This includes controlling physical devices in order to implement specifically logical NE functions,and processing and forwardin

224、g control commands from the cloud layer to the edge layer.The computing architecture can consist of the cloud layer,computing and network integration system(including the edge layer),and the device layer.The details are as follows:The device layer is where user tasks are initiated and generally cons

225、ists of PCs,mobile phones,wearable devices,and IoT devices.Advanced devices,such as powerful smart phones and 24 embedded single-board computers,can implement the functions of local gateways and simple networking.The edge layer is the bottom layer of the computing and network integration system.This

226、 layer receives tasks initiated by the device layer,and performs local computing or forwards tasks to higher-layer nodes according to scheduling policies.The edge nodes can be devices deployed in the environment.It is generally considered that real-time control and operation functions should be impl

227、emented near edge devices.That is,operations that have high requirements for low latency are generally performed at the edge layer of the computing and network integration system.The cloud layer is the top layer in the computing and network integration system.This layer provides application-layer in

228、terfaces,control commands,and domain-specific applications for devices.It integrates NFs and back-end applications to implement interworking between the computing and network integration system and the upper application layer.The preceding deployment architecture of the multi-layer computing and net

229、work integration system enables innovative services and applications that integrate computing and communication.In particular,computing power can be migrated from the network center/upper layer(i.e.,cloud layer and computing and network integration)to a location closer to users(i.e.,edge layer and d

230、evice layer),to achieve the following gains:Minimized latency:Some service applications involve only basic control loops.In other words,certain actions need to be triggered under certain time conditions.However,these control loops are generally time-sensitive.Performing analysis and quickly making d

231、ecisions locally or close to data sources can greatly reduce the latency of such control loops because the remote cloud layer may not be able to process data in real time and may have additional transmission latency.Higher reliability:Networks use sensor data to protect public safety or control crit

232、ical infrastructure.Uplinks to the remote cloud layer are subject to damage.Therefore,it is of great significance to consider local processing as an alternative,or to use local processing only.For example,edge computing can be used for local processing to improve service security and reliability in

233、an industrial control loop or an emergency response system.Privacy issue addressing:Some user data is sensitive or cannot be stored outside specific geographic boundaries according to laws.While cloud service providers are often considered trusted,users have no way of knowing and controlling where t

234、heir data is stored and who can access it.Local gateways and edge nodes are under the control of local operation personnel,providing better trust.Bandwidth saving:As the uplink bandwidth on the network is often limited,it is not always feasible to transfer large amounts of data from edge devices to

235、cloud servers.Processing or preprocessing data locally,and then transmitting the aggregated and filtered data to the cloud can significantly save uplink bandwidth.Cloud-edge-device collaboration:The system can perform cloud-edge-device collaborative scheduling based on service logic and actual resou

236、rces to provision services.The collaboration modes must include at least device-device collaboration,cloud-edge collaboration,and device-cloud collaboration.In production and daily life scenarios,there is no absolute division among cloud,edge,and device.The boundaries among them may be blurred and m

237、odified based on service content.In the deployment architecture of the multi-layer computing and network integration system,computing resources at all levels on the network can be systematically managed and effectively scheduled,and can be independently scheduled or jointly scheduled with communicat

238、ion resources.Therefore,different requirements of user services/applications on computing and communication performance can be met to the maximum extent,and user experience and resource utilization can be greatly improved.25 4.3.2 Algorithm Architecture Communication networks,especially mobile commu

239、nication networks,are characterized by an extremely complex network system,distributed deployment mode,ultra-large-scale node distribution,and requirements for the system to run stably.Therefore,the requirements and impacts of these communication network features on AI algorithms and algorithm archi

240、tectures should be fully considered during the design of the native AI algorithm architecture for new 6G networks.AI algorithms should be designed from the perspective of the new 6G mobile system architecture in order to implement a network AI system.The native AI algorithm architecture of new 6G ne

241、tworks should meet at least the following requirements:Stability:Carrier-grade mobile networks need to run stably,and this poses high requirements for the stability of AI algorithms in operations and maintenance.The AI algorithm architecture must have a stable structure and systematic design,and pro

242、vide strong fault tolerance capabilities.Real-time management and control:The native AI algorithm architecture must support intelligent real-time monitoring,management,and control of network information.During the design of the AI algorithm architecture,the real-time performance requirements and rea

243、l-time service assurance of telecom networks must be fully considered.Security and privacy protection:The AI algorithm architecture must natively support security and provide privacy protection for parameters transferred between algorithm modules and computing devices.Hybrid and layered design:Telec

244、om mobile networks are deployed in a hybrid manner(distributed+centralized)at different layers.At the architecture level,native AI algorithms should support distributed deployment and operations,as well as hybrid,layered,flexible deployment and operations,to meet the requirements of different layers

245、.In-network computing:As in-depth research on computing and network integration is continuously being conducted,in-network computing will become one of the key features of new 6G networks.Therefore,this network feature should be fully considered during the design of native AI algorithms,and the requ

246、irements for in-depth computing-network coordination should be reflected in the architecture design.Distribution of models and data:Due to differences in the capabilities and deployment modes of computing devices on the network,the AI models of communication networks,network data division,and distri

247、buted training and storage need to be fully considered during the design of the distributed algorithm architecture.In this way,deployment requirements and in-network computing requirements can be better met.Model exchange:Distributed or hybrid AI architectures require computing model interaction acr

248、oss layers,functions,and regions.Therefore,the definition and design of AI model transmission(exchange)modes also determine the performance of distributed AI algorithms/models.The AI algorithm architecture must be able to support synchronous,asynchronous,and hybrid model exchange based on algorithms

249、 and scenarios.Model-and data-driven:The traditional telecom mobile network design driven by knowledge models has distinct advantages in many aspects and can compensate for the weak generalizability and lack of explainability of data-driven ML algorithms.Therefore,the AI algorithm architecture shoul

250、d natively meet data-and model-driven requirements.Heterogeneous data:The particularity and heterogeneity of wireless data also need to be considered during the design of the AI algorithm architecture.Communication efficiency:The requirements for communication efficiency need to be fully considered

251、during the design of the native AI algorithm architecture in order to compress model updates and improve communication scheduling.In addition,the convergence and generalizability of AI algorithms/models have a huge impact on communication efficiency,and therefore need to be considered during the des

252、ign of the AI algorithm architecture.26 Evolvable and self-evolvable:The AI algorithm architecture should be highly evolvable to cope with network technology development and new service changes.4.3.3 Data Architecture Data is the foundation of AI algorithms/models.For new 6G networks that need to su

253、pport native AI capabilities and services,AI/ML-related data management functions can be considered as basic control functions that are comparable to mobility management and session/task management in importance.Therefore,the logical architecture of data management is crucial to the architecture of

254、new 6G networks.It should mainly be reflected in the organization and orchestration of data function entities.Figure 4-3 Example logical architecture of 6G data management(1)Figure 4-3 shows a possible logical architecture of 6G data management,where the following logical function entities are inclu

255、ded:Data exposure function(DEF):exposes data as a service(DaaS)of new 6G networks to application functions(AFs).The DEF performs privacy protection and aggregation on the 6G network data,and provides the processed data as data services for external or third-party AFs or AFs inside the system based o

256、n the data service requirements and/or service level agreements(SLAs)provided by the AFs.Data collection function(DCF):collects network,service,and device data in the new 6G mobile system for DAF entities.Data analytics function(DAF):trains AI/ML models based on the data collected by the DCF and use

257、s AI/ML algorithms to obtain analysis result information related to networks,services,and devices.The DAF can be further divided into two subfunctions:MTLF and AnLF.Data storage function(DSF):stores AI/ML data.Data security management function(DSMF):provides privacy protection and security control f

258、or AI/ML data.The DSMF interacts with the user subscription and authorization management(USAM)function and other data management functions(DCF,DAF,DSF,and DEF)to check the authorization,27 modification,and revocation of user data processing(e.g.,collection,analysis,and storage)and perform event noti

259、fication if required,to control the utilization and processing of user data by other data management functions.For example,after user data utilization authorization is revoked,the collection and analysis operations on the user data are terminated,and the stored data is deleted.Data resource orchestr

260、ation function(DROF):manages(e.g.,allocates,schedules,and deletes)resources related to AI/ML data processing.The DROF can also dynamically and intelligently orchestrate and manage data processing resources based on AI/ML algorithms(e.g.,deep reinforcement learning).Figure 4-4 Example logical archite

261、cture of 6G data management(2)Figure 4-4 shows another possible logical architecture of 6G data management.Similar to the architecture shown in Figure 4-3,the architecture shown in Figure 4-4 also includes the DEF,DSMF,and DROF entities.The only difference is that the architecture shown in Figure 4-

262、4 defines the data management function(DMF),which contains the DCF,DAF,and/or DSF.Therefore,the example architecture shown in Figure 4-4 can be regarded as a variant of the one shown in Figure 4-3.The function entities related to data management in Figures 4-3 and 4-4 can be deployed in a centralize

263、d or distributed manner as follows:The DEF entity can be deployed on the cloud or network edges,and be used to provide data services for application servers/AFs on the cloud or local application servers/AFs on network edges.The DMF can be deployed on the cloud or network edges,or be deployed on a 6G

264、 network in a distributed manner to support ubiquitous AI capabilities on the network and multi-point collaboration with other DMF entities The DCF,DAF,and DSF can also be deployed on the cloud or network edges,or be deployed on a 6G network in a distributed manner.The DSMF can be deployed on the cl

265、oud or network edges,or be deployed on a 6G network in a distributed manner to meet different privacy protection and security control requirements of AI/ML-related data in different regions of the network The DROF can be deployed on the cloud or network edges,or be deployed on a 6G network in a dist

266、ributed manner to meet real-time performance and localization requirements of data processing and resource management.28 The DEF,DSMF,DMF,DCF,DAF,and DSF are all control-plane function entities.The DEF and DSMF are generally deployed on the CN(or co-located with other CN function entities).The DMF,D

267、CF,DAF,and DSF can be deployed on the CN or RAN(or co-located with other CN or RAN function entities).The DROF is a management-plane function entity.In the preceding logical architectures of 6G data management,the types of data include but are not limited to the following:Input data for AI analysis

268、and inference,such as network data,service data,and device data.AI/ML model information,including model parameters/hyperparameters and model description file information(e.g.,applicable analysis types of the model,application conditions/status of the model).AI analysis result information,that is,ana

269、lysis result obtained by AI/ML model algorithms based on input data.Such information includes the predictive information about future behavior or status of networks or devices,or statistical judgment information about the current behavior or status.Cloud-edge collaboration:Data collection,processing

270、,and governance in a distributed collaboration manner are required.1)The cloud is responsible for providing a unified collection and scheduling portal,and the edge executes specific collection tasks.2)The cloud is responsible for providing a unified data processing portal and cooperating with the co

271、llaborative computing engine of the edge to schedule and control the collaborative computing logic.The edge is responsible for executing the push-down operators.3)The cloud is responsible for unified data governance portal provisioning,unified metadata registration,statistical analysis,and rule conf

272、iguration for different types of tables.The edge implements data governance.4.3.4 Other Architectures To implement 6G native AI,new 6G networks need to embrace a broader ecosystem than traditional communication networks,because it is difficult for one party to provide all the three AI elements(compu

273、ting power,algorithm,and data).In many AI scenarios,use cases need to be completed by means of multi-party collaboration.For example,new 6G networks can enable multiple enterprises to implement trusted federated training without directly sharing their IoT device data.Therefore,to provide native AI c

274、apabilities and services,new 6G networks need to natively support multi-party collaboration and participation mechanisms in terms of the computing,algorithm,and data architectures.Such mechanisms must enable multiple parties to participate in specific AI training,verification,or inference processes

275、in a simple,transparent,fair,and trusted manner,including but not limited to the following aspects:Computing power:Heterogeneous computing resources of new 6G networks can be constructed,shared,and maintained by multiple parties.Users and customers can,in a fair manner,transparently share their idle

276、 computing resources based on the 6G native AI architecture.Algorithm:AI algorithms in new 6G networks,including network-based AI and third-party AI applications,can come from various sources.The native AI architecture of 6G needs to support efficient management and operation of related algorithm li

277、braries and trusted verification of multi-party AI algorithms.Multiple parties can participate in different types of distributed AI training and decision-making.Data:AI data sources of new 6G networks can be any third party.The 6G native AI architecture needs to natively support the security of mult

278、i-party data and provide privacy protection.For example,when a user participates in a 6G federated training task,gradient synchronization in the training process should not result in the disclosure of user privacy.29 5 Summary of Technical Requirements and Principles This white paper briefly reviews

279、 the history and status quo of applications integrating AI and 5G mobile systems,and analyzes the foundation and transformation trends of native integration applications for AI and the new 6G mobile system in the future.During such a challenging evolution,the industry must first have a complete and

280、unified understanding of the technical requirements of 6G native AI in terms of capabilities,services,and architectures,and then build and extend native AI capabilities,services,and architectures.The implementation of 6G native AI involves all the three AI elements(computing power,algorithm,and data

281、)in terms of technical requirements for capabilities,services,and architectures.These technical impacts and requirements greatly differ from centralized AI systems represented by cloud AI today.In general,native AI is ubiquitous,distributed,local,collaborative,robust,private,and open.In addition,to

282、support differentiated and customized AI capabilities and service applications in the future,6G native AI needs to support the QoAIS assurance and evaluation mechanism in order to support new AI business models.6G native AI needs to support and build new forms of business that are broader than those

283、 that are built using traditional ICT networks.It must not only have more extensive,secure,and trusted providers of computing power,algorithms,and data,but also empower and serve more industry customers and users in a wider,deeper,and more profound manner.30 References 1 IMT-2030(6G)Network Working

284、Group,6G network architecture vision and key technology outlook white paper R,IMT-2030 6G Seminar in China,2021.2 China Mobile Research Institute,et al.,Research on 5G wireless network intelligence R,CCSA TC5 WG9,2021.3 3GPP,3GPP TS 22.261:Service requirements for the 5G system(Release 17)S,2021.4 3

285、GPP,3GPP TS 23.288:Architecture enhancements for 5G System(5GS)to support network data analytics services(Release 17)S,2021.5 3GPP,3GPP TS 28.104:Management and orchestration;Management data analytics(Release 17)S,2021.6 3GPP,3GPP TR 37.817:Study on enhancement for data collection for NR and ENDC(Re

286、lease 17)S,2021.7 Ye Ouyang,Lilei Wang,Aidong Yang,et al.,The next decade of communication artificial intelligence J,Telecommunications Science,2021,37(3):36.8 Qiang Cheng and Zishan Liu,Development status and future prospect of telecommunication network intelligence J,Information and Communications

287、 Technology and Policy,2020(9):7.9 China Unicom,China Unicom unified identification and service of heterogeneous computing power white paper R,2021 10 China Unicom,China Unicom cloud-network synergy to in-network computing technology evolution white paper R,2021 11 ZTE,ZTE autonomous evolving networ

288、k white paper R,2020.12 Mohammad Sultan Mahmud,Joshua Zhexue Huang,et al.,A survey of data partitioning and sampling methods to support big data analysis J,Big Data Mining and Analytics,vol.3,issue:2,pp 85101,2020.13 Jinxiang Peng and Li Zhang.Big Data Processing Technology and Application M,Tsinghu

289、a University Press,2020.14 Yongyong Ge,A survey on big data in the age of artificial intelligence C,pp.7277,2019 6th International Conference on Information,Cybernetics,and Computational Social Systems(ICCSS).15 Jinkang Zhu,Chen Gong,Sihai Zhang,Ming Zhao,and Wuyang Zhou,Foundation study on wireless

290、 big data:Concept,mining,learning and practices J,China Communications,vol.15,issue:12,pp.115,2018.16 Yuanwei Liu,Suzhi Bi,Zhiyuan Shi,and Lajos Hanzo,When machine learning meets big data a wireless communication perspective J,IEEE Vehicular Technology Magazine,vol.15,issue:1,pp.6372,2020.17 Guangyi

291、 Liu,Juan Deng,Qingbi Zheng,Gang Li,Xin Sun,and Yuhong Huang,6G native intelligence:Technical challenges,architectures,and key features J,Mobile Communications,2021,45(04):6878.18 Y.Yang,Multi-tier computing networks for intelligent IoT,Nat.Electron.,vol.2,pp.45,Jan.2019.31 Acronyms and Abbreviation

292、s 3GPP 3rd Generation Partnership Project 5GC 5G Core 5GS 5G System ADRF analytic data repository function AI artificial intelligence AI4NET AI for Network AN access network AnLF analytic logic function CN core network CSI channel state information DaaS data as a service DAF data analytics function

293、DCCF data collection control function DCF data collection function DEF data exposure function DL deep learning DMF data management function DROF data resource orchestration function DSF data storage function DSMF data security management function MCC mobile cloud computing MDT Minimization of Drive

294、Tests MEC mobile edge computing ML machine learning mMTC massive machine-type communications MTLF model training logic function NET4AI Network for AI NF network function NWDAF network data analytic function OAM operation,administration,and maintenance 32 ODICT operation data information communicatio

295、n technology QoAIS quality of artificial intelligence service QoE quality of experience QoS quality of service RAN radio access network RL reinforcement learning SBA service-based architecture SLA service level agreement ToB to business ToC to consumer URLLC ultra-reliable low-latency communication URSP UE route selection policy USAM user subscription and authorization management function XR extended reality 33