Contents

I. Introduction ..............................................................3
II. Motivation and Development History of IDAN ................................. 4
2.1 Research Background and Motivation ............................... 5
2.1.1 Challenge 1: Diverse and Changing Service Requirements of Users
..................................................................... 5
2.1.2 Challenge 2: Heterogeneity and Complexity of Intelligent Networks
..................................................................... 6
2.1.3 Challenge 3: Personalized and Customized On-Demand Services 6
2.2 Development History of IDAN ...................................... 6
III. Conception Meaning and Advantages of IDAN ................................ 8
3.1 Concept of IDAN .................................................. 9
3.2 Techniques of IDAN ............................................... 9
3.3 Advantages of IDAN .............................................. 12
IV. Architecture and Experimental Validation of IDAN ............................ 13
4.1 Hierarchical Operation Architecture of IDAN ..................... 13
4.2 Technical Implementation Architecture of IDAN ................... 14
4.3 End-to-End Implementation of IDAN ............................... 16
4.4 Experimental Validation of the Intent-Driven Optical Network
Architecture ............................................................ 17
4.4.1 Design and Optimization of the Deployment Method .......... 17
4.4.2 Operation and Maintenance Methods and Policies ............ 20
4.4.3 Integration Methods Compatible with the Existing Network .. 21
V. Key Technologies Involved in the Full-lifecycle of IDAN ........................ 24
5.1 Top-Down Intent Fulfillment ..................................... 25
5.1.1 Intent Modeling ........................................... 25
5.1.2 Intent Understanding ...................................... 26
5.1.3 Policy Verification ....................................... 28
5.2 Intent Realization Knowledge Management ......................... 30
VI. Application Cases of IDAN ............................................... 32
6.1 Application Cases - Intent-Driven CLL ........................... 32
6.2 Application Cases - China Unicom ................................ 36
6.2.1 Intent-driven Autonomous Orchestration Case ............... 36
6.2.2 Intent-Driven Autonomous Energy Saving .................... 37
6.2.3 Automatic Drive Test Intelligent Evaluation Solution ...... 38
6.3 Application Cases - China Mobile ................................ 39
6.3.1 Intent-driven Hierarchical Service Assurance Case ......... 39

1 / 56
 6.3.2 Intent-driven Sensing Certainty Assurance Case ............ 41
6.3.3 Autonomous Intent O&M Assurance High-quality Industry-specific
Leased Lines ........................................................ 43
6.3.4 Implementation of General Intent Solutions via ONAP ....... 45
6.4 Application Cases - Emergency Communication Network ............. 47
6.4.1 Case Analysis of Call on Session Initiation Protocol (SIP) 47
6.4.2 Integrated Space-Air-Ground Communication Network for Emergency
Rescue .............................................................. 48
VII. Technical and Application Challenges of IDAN .............................. 49
7.1 Technical Bottleneck Analysis ................................... 49
7.1.1 Uncertainty of Global Information ......................... 50
7.1.2 Variability of Dynamic Intent ............................. 50
7.1.3 Full-Lifecycle Closed Loop ................................ 50
7.2 Aspects to Explore in the Future ................................ 51
7.2.1 Bottom-Up Intent Assurance ................................ 51
7.2.2 Large Model of Intent-Driven Network ...................... 52
7.2.3 Intent-Driven Cross-Domain Autonomy ....................... 53
Abbreviations ............................................................. 54
References ............................................................... 55
Participants ...............................................................56

2 / 56
 I. Introduction
In response to the on-demand services in all scenarios, the future networks are expected to
allow users to enjoy services, change networks and share resources as desired. In the future, all
scenarios will coexist in the network, and there will be many global network parameters, diverse
resource conflicts, highly complicated network configuration, and mismatched delivery. Therefore,
it is necessary to enable intelligent management and control of the network based on service intent.
Intelligent networks are the product of deep integration of networks and artificial intelligence (AI).
Related new network concepts include Intent-Based Network (IBN), Intent-Driven Network
(IDN), Autonomous Driving Network (ADN) and Autonomous Network (AN), etc. The main
promoters include standards organizations such as IETF, ONF, TMForum, and CCSA, ICT
equipment manufacturers such as Cisco and Huawei, and telecommunication operators such as
China Mobile and China Unicom. This white paper believes that the IDN is the key technology to
realize fully advanced autonomous networks in the future. The intent-driven autonomous network
(IDAN) is therefore defined herein, aiming to explore the concepts, architectures, technologies,
and use cases, and summarize future work.
IDAN has become a consensus in the industry. Various standards organizations, industry
organizations, open source communities, and other players in the communication industry are
actively promoting the industrial application layout of autonomous networks, covering all fields of
the industry, such as communication operators and communication equipment manufacturers. As
the technical research and development and commercial implementation of IDAN are accelerating
pace, the industry is booming.
Through top-down forward intent fulfillment and bottom-up reverse intent assurance, IDAN
establishes closed-loop life cycle management of a space-divided (from user space to digital space
to physical space) and multi-layer IDN. Based on the zero trust principle, a step-by-step feedback
method is adopted to continuously monitor the entire cycle for dynamic optimization.
By building automated, intelligent O&M capabilities throughout the network life cycle,
IDAN provides customers with the ultimate service experience of "zero wait, zero failure, and
zero contact", and create efficient O&M featuring "self-configuration, self-fix, and
self-optimization" for the production front line of the network. By deeply integrating AI with the
hardware, software, and systems of the communication network, IDAN facilitates agile innovation
of enablement services, intelligent network operation, and building of intelligence-endogenous
networks.
Building an intent-driven autonomous network is a complex system project that requires
complete top-level design, and unified definition of standards and technical processes. This white
paper analyzes the development trends of IDAN in a comprehensive manner, sorts out the
evolution trends of key technologies on IDAN, discusses the industrial application of autonomous
networks of operators and equipment vendors, and analyzes the technical challenges and
engineering problems of IDAN applications to guide subsequent research and application.

3 / 56
 II. Motivation and Development History of IDAN
Driven by the rapid development of the digital economy, the global telecommunications
industry is accelerating the transformation and upgrading towards automated, intelligent networks.
With the large-scale deployment and rapid commercialization of communication networks, the
rapidly developing digital economy has opened the door to the digital era. With the deep
integration of communication networks and AI, autonomous networks can provide consumers and
vertical industry users with innovative network services and digital experience [1].
Operators have gradually realized the importance of IDAN. China Mobile views the
autonomous network as a key trend in improving the quality and efficiency of communication
networks and facilitating the digital upgrade of the industry. With technologies such as IDN, it
helps enable automated deployment of sensing and control policies in the network and evolution
to advanced autonomous networks. China Unicom emphasizes that in order to meet the diverse
service needs of users, networks must have the characteristics of load response and service
customization in all scenarios. The introduction of the concept of intent can simplify network
operations, improve automation level, and meet service demands more effectively. China Telecom
proposed the concept of "intent-based network" at the 2019 SDN/NFV/AI Conference. Driven by
intents, IBN works to realize personalized network services and flexible allocation and sharing of
network resources. IBN aims to cover the entire society, industry, and ecosystem, and define an
intelligent management and control system for all scenarios and domains. Chinese operators are
leading the innovative practice of autonomous networks and promoting a consensus in the
industry across the world. Since 2021 when China Mobile first proposed the goal of reaching L4
in 2025, several companies including China Unicom and China Telecom have set relevant goals,
aiming to reach L4 level in high-value scenarios or across all networks.
Equipment vendors are launching intent-driven products and solutions. Huawei has proposed
in the industry the concept and grading standard of the autonomous driving network. It has defined
grading standard for autonomous driving in communication networks in terms of service
experience, labor-saving level, and complexity of the network environment. According to this
grading standard, L5, i.e. the fully autonomous network, which is the ultimate development goal
of telecommunications networks, is also the development goal of IDAN. ZTE has proposed the
"autonomous evolving network"(AEN) solution. By using ubiquitous AI to promote the
intelligentialization of the entire network, and adopting the hierarchical closed-loop principle to
build an intelligent network system at the network element level, single-domain level, and
cross-domain level, AEN aims to push the network system to gradually realize autonomous
operation, and, as driven by data, enable self-learning and self-evolution, and the intelligent
autonomous network system, so as to optimize the investment efficiency and O&M efficiency of
the network. ZTE has planned a comprehensive service process covering "planning, building,
maintenance, optimization, and operation" and more than 60 intelligent application scenarios,
including single-domain intelligent scenarios in various dedicated domains such as wireless,
bearer, and core networks, as well as scenarios of end-to-end digital operation across dedicated
domains.
Many universities are exploring breakthroughs in single technical points on autonomous
networks. Beijing University of Posts and Telecommunications has conducted research on
knowledge-defined intent-driven networks. Xidian University is trying to achieve a closed-loop
4 / 56
intent life cycle following the two technical paths: top-down intent fulfillment and bottom-up
intent assurance. University of Electronic Science and Technology of China has deeply explored
intent representation and the classification and modeling technologies, and applied the
intent-driven concept to emergency communications.
In order to seize the opportunity first, enterprises have been constantly making adjustments
and changes. In order to benefit from the market opportunities brought by key technologies in
future networks, cloud computing, and edge computing, operators must also make active changes
and devote themselves in. In the future service scenarios, not only networks with low latency, high
reliability, and support for massive device connections are required, but operators are also
expected to provide additional services, including massive data analysis, image recognition, and
precise positioning. Research activities on intent-driven autonomous networks are popular in
universities, showing the huge and unstoppable development potential of this field.

2.1 Research Background and Motivation

The development of the global digital economy is moving towards a new stage of "deep
application, standardized development, and inclusive sharing". The digital transformation is being
advanced in various industries, expanding from office and marketing service scenarios to core
production and manufacturing scenarios, shifting from efficiency change to value change, and
extending from internal digital transformation within the enterprise to collaborative transformation
along the industrial chain and value chain. Against this backdrop, the need for 5G applications and
cloud-based big data-driven intelligence has been more common and urgent among governments
and enterprises.
In the past ten years, China has achieved remarkable achievements in the development of the
digital economy. In 2022, China's digital economy registered a GDP of CNY 50.2 trillion, with a
year-on-year nominal growth of 10.3%, significantly higher than the nominal GDP growth rate in
the same period for 11 consecutive years. The digital economy accounts for 41.5% of GDP,
equivalent to the proportion of the secondary industry in the national economy. With its overall
scale ranking second in the world, the digital economy in China is increasingly playing a leading
and supporting role in economic and social development. 2023 witness the emergence of many
new digital services, such as computing-network integration, Internet of Vehicles, and HDICT,
provided by communication operators, as well as new business models, such as network as a
service, service customization, and project delivery. All this places higher requirements on
network infrastructure and automated intelligent operation.
Driven by the rapid development of the digital economy, the global telecommunications
industry is accelerating the transformation and upgrading towards automated, intelligent networks.
Against this backdrop, autonomous networks emerge. In order to meet the diverse digital needs of
vertical industries and consumer lives, the autonomous network framework and its key features
are proposed. However, the evolution of network technology and the improvement in O&M
management have brought many challenges.

2.1.1 Challenge 1: Diverse and Changing Service Requirements of Users

With the development of science and technology, various industries have increasingly urgent
needs for technological innovation. For example, the development of new energy, AI, big data and
other fields has created opportunities for transformation and upgrading across traditional
5 / 56
industries. User demands for network services are also changing faster and faster. In pursuit of
personalized and customized consumer experiences, users have more diverse demands for
products and services. Operator network services must have the agility to capture user intent,
intelligently and flexibly adjust resource allocation and operation policies based on user intent,
and improve product and service quality to meet the constantly changing needs of users and
enhance user experience. With the saturation of the individual user market, the digitization of
vertical industries and that of consumer lives have become the highly-expected potential growth
points, and more diverse needs have been proposed for connectivity, bandwidth, latency, reliability,
and other aspects of network services.

2.1.2 Challenge 2: Heterogeneity and Complexity of Intelligent Networks

With the dramatic increase in the number of devices connected to the network, the network
are constantly expanding. This will bring huge challenges to network management, e.g. how to
ensure reasonable allocation of network resources, how to improve network performance. In
addition, intelligent heterogeneous networks in the future will cover multiple access technologies,
multiple application scenarios, and deeply integrate with various industries and fields. This will
make the network architecture more complex and require better network management. How to
simplify network management and improve O&M efficiency will be one of the major challenges.

2.1.3 Challenge 3: Personalized and Customized On-Demand Services

On-demand services in all scenarios aim to provide users with personalized and customized
network services to meet the needs of different scenarios. With AI technology, intent-driven
intelligent policies can better sense user needs, enable intelligent scheduling and optimization of
network resources, and thus improve network performance and user experience. The integration of
on-demand services in all scenarios and intelligent policies can better meet users' network needs in
different scenarios and improve network resource utilization. The autonomous network has a
network system with autonomous learning, adaptive adjustment and self-optimization. It can
automatically adjust network parameters based on user behavior and network conditions to
reasonably allocate and optimize network resources. The IDN drives the adjustment and
optimization of network behavior with user needs and intent at the core. The integration of
autonomous networks and IDN brings many advantages, including improved network
performance, better user experience, greater network security protection, lower management cost,
and adaptability to the development requirements in future networks.
To sum up, in order to meet diverse service needs of users, the introduction of intent becomes
the key to solving problems in the future development of network load features and service
customization in all scenarios, so as to simplify network operations, improve automation level,
and better meet service needs.

2.2 Development History of IDAN

As many domestic and foreign standards organizations, including TM Forum, 3GPP, ITU-T,
ETSI, and CCSA, have been carrying out the work related to autonomous networks, enterprises
are actively promoting the standardized layout of autonomous networks to compete for a voice in
the industry [5-7].
In 2015, the Northbound Interface Working Group of the Open Networking Foundation
6 / 56
published the Intent-Based Networking White Paper. Gartner, Cisco, and Huawei have
successively proposed the definition of IDN, clarifying the basic components of IDN, such as
intent translation, intent verification, automatic implementation, and holographic sensing [7][10].
As a brand-new network model that analyzes user intent, IDN translates the intent into
corresponding network policies, and ultimately realizes automated deployment of network sensing
and control policies, and thus the network is evolved from a static resource system into a dynamic
system that can meet business goals continuously.
In 2015, ONF released a standard proposal titled Intent: Don't Tell Me What to Do! Tell Me
What You Want: In intent-based networks, intelligent software decides how to translate intent into
configurations for specific infrastructure so that the network can operate in the desired manner.
In 2017, Cisco proposed in the white paper titled Moving towards Intent-Based Networks:
The network team can use concise words to describe the work they want to accomplish, and then
the network can translate the intent into numerous policies that will make appropriate
configuration and setting changes in complex, heterogeneous environments with the automated
functions.
In 2017, the Network Management Research Group (NMRG) of the International Internet
Research Task Force IRTF launched the intent research, focusing on the analysis of concept
definitions, main principles, intent classification, and life cycle management, and released RFC
9315 and RFC 9316, which define intent as the abstract, high-level policy for operating networks
[5].
In 2018, the IDMS_MN project initiated by 3GPP SA5 described the concept and scenarios
of intent-driven management in terms of the mobile communication network management layer,
and recommended standardizing intent expression in the specification stage. 3GPP released the TS
28.312 standard in R18. In 2024, the FS_IDMS_MN_Ph3 project was launched in R19 to conduct
further research on enhanced intent-driven management services for mobile networks.
In 2019, ETSI launched the comprehensive research on intent standards. The first released
ETSI GS ENI 005 defines the ENI system policy management model, which supports declarative
policies, imperative policies and intent policies. ETSI GR ENI 013, released in 2022, focuses on
intent information model standardization, and carries out gap analysis based on the current status
of the industry. ETSI IFA 050, launched at the same year, mainly studies intent-driven interfaces
and model standards. ETSI SOL 021, launched in 2023, mainly studies the RESTful protocol and
data model specifications regarding the requirements on intent management service interface.
The autonomous network, jointly proposed by TM Forum and industry partners in 2019, aims
to lead the automated and intelligent transformation of network infrastructure and operation
systems. After more than four years of development, systematic concepts, standards,
implementation methods, and application cases of the autonomous network have formed, with
remarkable achievements in industry consensus, standard formulation, and practice and
deployment [7].
(1) Industry consensus: Zero-X/Self-X vision, L0-L5 grading, three-layer and
four-closed-loop architecture, single-domain autonomy and cross-domain collaboration,
intent-driven and full stack AI and many other concepts have become widely-accepted industry
consensus;
(2) Standard formulation: 9 major standards organizations, including TM Forum, 3GPP,
7 / 56
CCSA, and ETSI, have been focusing on 5 major standard directions, and have initiated/published
more than 80 standards/research subjects in total. In addition, they rely on the M-SDO Initiative,
involving multiple standards organizations, to ensure homogeneous architectures and unified
standards;
(3) Practice and deployment: 14 leading operators, including China Mobile, China Telecom,
China Unicom, Deutsche Telekom, and Vodafone, have included autonomous networks into their
group strategies. Driven by commercial value and operational effectiveness, they aim to conduct
hierarchical assessment and autonomous capability planning and building in an iterative way with
the goal of reaching L4 autonomy between 2025 and 2027.
In 2021, TMF officially launched the work on standards related to cross-layer and
cross-domain general intent. In the same year, it delivered IG1253 and a series of intent-related
specifications, and conducted in-depth research on intent modeling, intent information models,
intent life cycle management, interfaces and intent use cases [4]. In 2022, it launched research on
the TR290 general intent model, TR291 extended intent model, and TR292 intent management
ontology. And it published two research reports on the development of intent in autonomous
networks: Autonomous Networks: Empowering Digital Transformation - From Strategy to
Implementation in 2021 and Intent in Autonomous Networks in 2022 [8].
The NFV special project group SP1 of CCSA released a series of standards for technical
requirements for network functions virtualization (NFV) management and orchestration in 2021,
stipulating the overall architecture and functional requirements of intent management, intent
management information models and intent management interfaces in the technical requirements
for NFV management and orchestration. The Network Management Technical Committee TC7 of
CCSA launched technical research on intent management in autonomous networks in June 2022,
focusing on research and discussion on intent management architectures and related reference
points, intent expression models, intent life cycle management, etc. And it officially launched the
formulation of standards for technical requirements for intent management in autonomous
networks in June 2023 [9].
During the above-mentioned standardization research process, various standards
organizations and manufacturers have been actively devoted in the research on IDN and proposed
their solutions. IDNs are gradually developing from theory to practice, and are constantly being
improved and optimized in practice, showing great potential in network management, automated
deployment, and optimized network resource allocation. However, the current research and
practice of IDNs are still facing some challenges and problems. For example, how to accurately
understand and translate user intent, how to ensure the accuracy of intent verification and
implementation, and how to enable comprehensive sensing and dynamic adjustment of the
network. All these problems have to be studied and solved in future research and practice. In
addition, autonomous networks, as an important application scenario of IDNs, are getting more
and more attention. With the introduction of the concept of intent, autonomous networks enables
automated, intelligent, and dynamic networks, thereby better meeting the needs of users and
business development. Therefore, how to achieve effective management and application of intent
in autonomous networks will be an important subject in the future research.

III. Conception Meaning and Advantages of IDAN

8 / 56
 Based on the above-mentioned existing standardization research results, there is an urgent
need for a consensus on the definition of intent in autonomous networks, and the definition,
architecture and key technologies of IDNs. And the theoretical and technical system of IDNs
needs to be improved and optimized to better promote the development and progress of the
network. This chapter introduces the basic concepts of IDAN, including the overall framework,
core ideas, and key capabilities.

3.1 Concept of IDAN

As defined in TMF IG1253 [4], intent is the formal specification of all expectations,
including the requirements, goals and constraints for technical systems. The intent owner is the
creator of the intent and is responsible for the management of the intent life cycle. The intent
processor is the receiver of an intent, and is responsible for intent realization and satisfaction, and
the management of the life cycle of the intent instance.
IDAN is an advanced version of IDN. It emphasizes that the network no longer passively
receives service instructions, but tries to actively understand the intent of human administrators to
anticipate opportunities and act according to the intent as far as possible, and takes into account
both the accuracy of the intent and network optimization goals. In addition, it can dynamically
adjust network parameters to improve network performance according to the network environment
and user needs. According to the autonomous network grading standards, at the L4 (advanced
autonomous network) stage, IDAN are implemented by introducing intent. IDAN's meaning
covers the following aspects:
With user intent at the core: User intent is the main consideration in the design and operation
of network systems. Network resources are automatically adjusted by analyzing and understanding
user needs. It also takes time to provide optimized services. Not all intent is of equal priority, so it
should configure a priority for each intent and reserve that of high priority.
Automated and intelligent: With advanced automation and intelligence technologies, such as
artificial intelligence and machine learning, network resources can be optimally allocated, and
automatic configuration of intent priority, optimization, and failure recovery can be enabled. These
technologies can learn, sense and make decisions autonomously, thereby improving the
adaptability of the network.
Zero wait, zero contact, and zero failure: With precise resource control and automated O&M
capabilities, real-time service provisioning, ready-to-use and zero-wait experience can be realized;
with end-to-end network monitoring and intelligent configuration failure recovery, zero-failure
experience can be realized; with open network data and capabilities and self-services, zero-contact
experience can be realized.
Network autonomy in all scenarios: Building systematic capabilities based on the four
management layers of network elements, resources, service, and business, enables network
autonomy in all scenarios. By building systematic capabilities at different layers, collaboration and
closed-loop control between different management levels can be achieved, which boosts external
business growth and improves internal efficiency.

3.2 Techniques of IDAN

An autonomous network is a network infrastructure, operation management system and
service system that integrate intelligence and automation [2]. At present, the automated O&M
9 / 56
capabilities of autonomous networks rely largely on preset empirical rules carefully devised by
experts. However, these rule-based automated policies may face many challenges in real-world
applications. First, the design of rules cannot cover all scenarios, resulting in poor performances in
specific or complex environments. Secondly, parameter configuration is usually static, and cannot
adapt to instant changes in the network environment, reducing system adaptability and response
speed. In addition, parameter configurations are often divided into several fixed ranges, and
cannot be refined according to actual needs, affecting resource utilization efficiency. Finally,
intervention of experts is required to identify and resolve potential conflicts between rules,
increasing operational complexity and cost. In response to these problems, autonomous networks
need to use more intelligent and flexible technologies to adapt to the constantly changing network
environment and needs. This is a key technical research direction for implementing IDAN.
(1) External input intent and endogenous intent
According to the way of generation, intent in IDAN can be divided into human-generated
intent and automatically generated intent. Human-generated intent is called extrinsic intent, and
automatically generated intent is called intrinsic intent. The key technology mainly involved in
extrinsic intent is intent understanding. Intrinsic intent is endogenous intent that is usually evolved
to better accomplish human intent, and the key technology involved is intent prediction.
"Autonomous" is reflected in the process of intent understanding and prediction. In intent
understanding, the system must have the ability to analyze and understand user input, which may
involve natural language processing, image recognition, speech recognition and other technologies.
In intent prediction, the system needs to use data analysis, model training, and other methods to
predict the possible next intent of the user or evolve secondary intent based on the user's current
intent to complete the current service. These operations enable the system to intelligently sense
and predict user intent, making the network serve users in a better and smarter way.
(2) Intent-based prediction technology
Intent prediction technology plays a unique role especially in emergency scenarios. In sudden
emergency scenarios such as natural disasters, man-made accidents, or infrastructure failures,
IDAN can automatically and quickly schedule network resources to deal with the incidents. Take
the prevention of and response to sudden forest fires as an example. Emergency disasters usually
require millisecond-level response. However, due to the complicated network integrating air,
space, and ground, a large number of network configuration operations are required, which places
a huge burden on human operators. Through IDAN, a large number of sensors and satellite
imagery can be heterogeneously accessed with the help of the intent-driven proxy to automatically
collect warning messages from the physical network. With efficient intent translation and rapid
decision-making, intent can be automatically and accurately issued to direct relevant personnel to
take emergency measures, and related resources across the entire network can be automatically
scheduled to provide immediate communication and coordination support. In addition, with the
help of big data analysis and machine learning, autonomous networks can predict the spread
direction of the fire, so as to predict the possible development trend of the disaster in advance,
generate various plans for human administrators, and help take more timely and effective
emergency measures.
(3) Intent-driven management service
The intent-driven management service is a key component to implement the intent-driven
10 / 56
autonomous function. In SDN-based implementation, it can be implemented on the SDN gateway
and interact with the intent translation function through the southbound interface to translate the
intent in natural language into that using the intent model. The intent-driven controller is
responsible for parsing and delivering the intent to the intent-driven management service, and the
proxy is responsible for the specific network configuration. The intent-driven proxy provides
access and interconnection support for end links, such as satellite links, self-organizing networks,
and cluster networks, through physical network interfaces.
(4) Intent priority
Network resources are always limited, and it takes time to schedule them. Various
communication services are not of equal priority. Therefore, important intent can be configured
with higher priority to prioritize key communication services by reserving network resources and
other means. The configuration of intent priority plays an important role in resolving intent
conflicts in emergency communications. The intent priority may be affected by the clarity of user
intent, context, urgency, system policy, user expectations, task complexity and other factors. Even
the same intent, if issued by different users at different times, its priority may not be the same.
(5) Intent pyramid model
The first stage is configuration detail management, represented by SNMP. This stage focuses
on the basic monitoring and configuration of underlying network devices. The second stage is
system policy management, represented by PBNM. This stage introduces more complex policies
to make network management more targeted and flexible. The last stage is service intent
management, represented by IDNM. This stage emphasizes the implementation of a higher level
of adaptive network management based on the user's service intent. During the evolution of the
three stages, as the abstraction level of the underlying facilities is gradually raised, the
complicated technical details are continuously screened out, finally implementing network
management based on the service intent of the highest abstraction level, forming mapping from
"intent" to "policy" to final "configuration" and providing more flexible and intelligent means for
network management. The pyramid model of “intent-policy-configuration” is shown in Figure 1.

11 / 56
 Figure 1 Schematic Diagram of the "Intent-Policy-Configuration" Pyramid Model

3.3 Advantages of IDAN

IDAN shows four highlights. First, it brings greater flexibility, intelligence and efficiency to
the business model, provides users with more innovative and valuable products and services,
offering enterprises new business opportunities and growth space. Second, by capturing data from
the entire network in real time and leveraging massive data analysis, in-depth insights, predictions
and knowledge extraction, the potential of data flow is released to facilitate data-based product
innovation. Third, it improves resource scheduling, automation and intelligence, effectively
reduces labor cost, and enhances the utilization efficiency of terminals, base stations and
computing centers, thus cutting operating cost for enterprises. Finally, IDAN has excellent risk
identification, prediction and emergency response capabilities, significantly reducing potential
risks and losses. By using automation and intelligent means to dynamically balance the usage
needs of the entire network, producing remarkable effect on energy saving and emission reduction
[3].
With its unique advantages, IDAN is reshaping the traditional network operation model [11].
IDAN has remarkable advantages in improving network performance, optimizing resource
utilization, and providing personalized services, which help promote the innovation and
development of network technology, mainly reflected in the following aspects:
(1) Improving network resource utilization: By monitoring network state and user needs in
real time, IDAN can intelligently adjust network resource allocation for efficient resource
utilization and waste reduction.
(2) Optimizing network performance: IDAN continuously optimizes network performance
based on advanced algorithms and models. This helps improve network stability and reduce
latency and packet loss, thus improving user experience.
(3) Automated and intelligent: With autonomous learning capabilities, IDAN can extract

12 / 56
valuable information from massive data, and automatically adjust network parameters and policies.
This offers more convenient network management and lower manual intervention cost.
(4) Personalized services: IDAN can provide users with personalized and customized
network services based on user behavior and preference. This helps improve user satisfaction and
loyalty, which is an important direction for business development.
(5) Open and scalable: With good openness and scalability, IDAN can be seamlessly
integrated with other network technologies and applications to adapt to constantly changing
network requirements. This provides strong support for network innovation and development.
(6) Coping with complex scenarios: IDAN can cope with diverse network environments and
complex scenarios, optimizing support for various applications. This helps increase network
adaptability and flexibility.

IV. Architecture and Experimental Validation of IDAN

4.1 Hierarchical Operation Architecture of IDAN

The deployment practice of multilateral access gateways and autonomous networks centers
around the core idea of "single-domain autonomy and cross-domain collaboration" to build
systematic capabilities in layers to achieve network autonomy in all scenarios. Its target
architecture includes four layers, i.e. network element management (dynamic sensing and
automatic optimization of equipment components and operation state, and opening up automated
operation capabilities), network management (automatically converting the upper-layer service
and application intent-driven controller into network behavior based on network management,
control and analysis capabilities, and, with the intra-domain closed-loop control capability of
southbound sensing/analysis/decision-making/execution, in combination with local intelligence,
continuously guaranteeing the SLA commitment of network connections or functions and thus
achieving closed-loop management of single-domain network autonomy), service management
(based on the end-to-end process of planning, building, maintenance, and optimization, in
combination with the three major capabilities of service collaboration, guarantee, and analysis,
enabling cross-vendor and cross-domain service layer autonomy and closed-loop management)
and business management (mainly targeting at autonomous network businesses, providing
customers, ecosystem and partners with business enablement and operation capabilities) from
bottom to top, as well as three closed loops, i.e. the resource closed loop (managing resources of a
single field, achieving single domain autonomy), business closed loop (business-oriented
end-to-end management across fields to achieve cross-domain collaboration) and user closed loop
(user and business management, including user information, operation, billing, and customer
service) [2]. The hierarchical architecture model reduces the complexity of the overall system.
Each layer can evolve independently and operate autonomously, and hide the domain
implementation technology, intra-domain operations and intra-domain functional details from the
interface and its consumer layer, as shown in Figure 2. Among them, the A-type reference point is
located between the hierarchical decoupled management architecture layers, which can serve as
the autonomous system northbound to provide the upper layer with the call and management entry
for intent processing capabilities. The K-type reference point is located between the centralized
autonomous node and the hierarchical decoupled management layer, providing the latter with the
call and management entry of centralized intelligent support capabilities required to achieve intent
13 / 56
management.
An intent overlay layer can be introduced at each layer of the autonomous network and
interact between layers to provide access and interconnection support for end links such as
satellite links, self-organizing networks, and cluster networks. The multilateral access gateway
provides access interfaces for heterogeneous network interfaces, and the intent-driven controller is
responsible for simplification and abstraction. In the intelligence stage of autonomous network
evolution, based on user-specified intent goals and combined with AI technology, the system
provides closed loop automation within layers. Heterogeneous network intent interface
interactions between each layer are controlled through the gateway, so that upper-layer service
calls can be implemented independent of the lower layer. This enables intent issuance and intent
coverage across layers and domains of the autonomous network, further reducing human
operations, improving management efficiency, and accelerating technical iteration.

Figure 2 IDAN Architecture

4.2 Technical Implementation Architecture of IDAN

Figure 3 shows a reference technical implementation architecture of IDAN, which adopts a
layered design. According to the functional abstraction level and service logic of each layer, the
network is divided into the infrastructure layer, service management layer and business
management layer from the bottom up. The distributed artificial intelligence (DAI) module spans
all layers and provides powerful support for each process.
The infrastructure layer, consisting of various hardware facilities, software systems, and data,
can be viewed as the cornerstone of IDAN. This layer senses and manages computing resources,
network resources, and storage resources across the network in a unified manner, and can
excellently manage and schedule computing and storage resources based on service needs. With
DAI, the infrastructure layer expands the depth and dimension of its own information sensing,
including resource sensing, performance sensing and failure sensing, providing reliable and
comprehensive decision-making input to the network management layer. In addition, through
analysis and decision-making at the data source, functions such as real-time uninterrupted service

14 / 56
response, intelligent adjustment of equipment energy consumption, and computing network failure
sensing and repairing can be enabled to improve the self-response, self-healing, and
self-optimization capabilities of the system.
Regarded as the brain of IDAN, the network management layer is responsible for
implementing system functions. This layer receives the state information and service intent of the
infrastructure layer through northbound and southbound interfaces respectively, and performs
analysis, decision-making and network control, including state awareness, resource scheduling,
computing power management, service orchestration, and failure analysis and self-healing. The
network management layer adopts the hierarchical progression of single-domain autonomy and
cross-domain collaboration, emphasizing automated, intelligent execution of sub-functional
modules of the system, while supporting adaptive evolution of resource state and user intent.
Cross-domain collaboration connects various autonomous domains, solves complicated problems
through multi-domain collaboration, and enables the automation of functional processes and
high-level intelligent closed-loop processing of services. The network management layer deeply
embeds AI into all levels of the computing network to improve the intelligent learning ability and
scenario adaptability of each functional module to ensure the quality of service of the computing
network in providing current and new services.

Figure 3 Design Framework of IDAN

The service application layer is used to open up user-oriented service capability and bear
abstract service functions. From the user's perspective, the service application layer can schedule
the service application to appropriate nodes based on user intent, achieving optimal resource
utilization and ensuring excellent user experience.
The DAI module forms a complete intelligent closed loop through data management,
learning training, intelligent distribution and continuous learning, providing comprehensive
intelligent services for the infrastructure layer, network management layer and business
application layer. As the intelligent capability management and knowledge unification center, the
15 / 56
DAI module is deeply integrated at all levels, and the entire life cycle of AI, including design and
training, inference verification, deployment and application, and iteration and optimization are all
deployed within IDAN, so that AI is born and served internally. By deeply mining the data
generated during the operation of the computing network, the DAI module collaboratively
integrates the different data, resources and functions between different network layers to generate
effective solutions in consideration of operation efficiency, automation level and service quality,
and other aspects of IDAN. At the same time, this module supports AI in continuous learning, with
the ability to adaptively evolve knowledge and perform knowledge fusion and inference to create
new knowledge.

4.3 End-to-End Implementation of IDAN

A complete intent processing process includes the following stages, i.e. the start stage,
evaluation stage, issuance stage, and implementation stage, as shown in Figure 4.
In the start stage, the intent owner incubates new intent objects by detecting how well the
requirement can be met to determine whether to define a new intent object or change the current
one. If there is no need to create a new intent object, the processing process ends.
In the evaluation stage, the intent owner and potential intent processor determine the feasible
implementation method for the intent object through investigation and negotiation, including the
selection of the intent processor and negotiation on intent object parameters, and verify and
evaluate the impact of the corresponding method. If it is deemed not feasible by the evaluation, the
processing process ends; If deemed feasible, the intent owner will determine the intent processor
and corresponding parameter information required to implement the intent object.
In the issuance stage, the intent owner issues the evaluated intent object to the intent
processor to request a new intent instance. If the intent processor accepts the intent object, the new
intent instance is created, and the intent life cycle enters the implementation stage. If the intent
processor does not accept the intent object, the creation of the new intent instance is failed, and the
processing process ends.

Figure 4 Schematic Diagram of the End-to-End Intent Processing Process

In the implementation stage, the intent processor operates its responsibility domain according
to the accepted intent to realize the desired goal of the intent, continuously ensures that the

16 / 56
corresponding expectations of the intent are met before the intent instance is deleted, and reports
the intent processing progress, execution status, the reason why the requirement cannot be met, etc.
to the intent owner as needed. When the intent processor receives the request to update the intent
instance, it determines the feasible method for the update operation through investigation and
negotiation, and verifies and evaluates the impact of the corresponding method. If the update is
not feasible, the processing process ends; if the update is feasible, the corresponding intent
instance will be updated. When the intent processor receives the request to delete the intent
instance, the processing process ends.
In an IDN, depending on different layers of business management, service management,
resource management, and network element management, the upper-layer intent management
function can serve as the intent owner, and the lower-layer intent management function can serve
as the intent processor. The top-down intent calling process is implemented between layers by
calling the intent interface, so that the user's end-to-end intent can be satisfied.

4.4 Experimental Validation of the Intent-Driven Optical Network

Architecture
Based on the IDAN framework, the intent-driven optical network platform (IDONP) is
developed, including the optimized design of deployment solutions from intent to policy, a fault
location mechanism for high-precision intent assurance, and a fast slice reconstruction method for
intent assurance.

4.4.1 Design and Optimization of the Deployment Method

This section explains the requirement analysis of IDAN, and two implementation methods
from intent to policy, including intelligent policy generation based on strong intent constraints
(PG-RL) and intelligent slicing policy generation based on intent constraints, for optimal
adaptation of intent to policy.
As the number of access users and service types increases, the traffic of optical networks
changes in stages and regionally. However, the cost of building and demolishing routes is very
high, and it is difficult to ensure real-time building and demolition. Therefore, it is necessary for
the network to be able to appropriately and automatically change on the existing network
infrastructure, which can be done through network automation or network orchestration, and is
required to: 1) automatically generate configuration policies under strong intent constraints to
meet the service requirements of the optical network; 2) establish a highly robust intent assurance
mechanism; 3) realize real-time adaptation of intelligent policies to the optical network
environment.
Technical principles: The mapping of intent to slicing policy implementation usually requires
configuring the connection topology (connecting to service nodes), exchanging data to update
content, and allocating sufficient computing, storage, and transmission resources to maintain a
certain QoS level. The optical network slicing can adapt to the flexible requirements of diverse
services and provide customized service guarantees. Network slicing allows infrastructure
providers (InP) to support heterogeneous services on a public platform (i.e. create a customized
slice for each service). And the slice can be dynamically scaled up/down to match any change in
the requirements of its services. The slice is created based on the requirements for the intent
translation result (e.g. latency, capacity, and reliability). In the case where there is a change in time
17 / 56
and/or space in the traffic aggregated in the slice, the InP can dynamically scale up/down the
provisioned slices to match the change in service requirements, thereby improving its resource
utilization efficiency. In addition, benefiting from the development of network functions
virtualization (NFV) and software-defined networking (SDN) technologies, network slicing has
been proposed as a key architecture technology in the intent-driven integrated optical-wireless
network. DRL learns from experience by interacting with the network, constantly searches for and
adjusts appropriate policies, and dynamically adjusts the resources allocated to each slice, thereby
maximizing resource utilization while guaranteeing the intent constraints.
Deployment method: The intent-driven integrated optical-wireless network architecture
mainly includes intent expression, translation, verification and deployment. In terms of intent form
translation, the intent flow of "natural language, intent primitive, executable policy, and reliable
configuration" is formed. In terms of hierarchical relationship, combined with the latest practices
of Open Daylight (ODL) and Open Network Operating System (ONOS), the intent-driven
integrated optical-wireless network mainly includes the service application layer, intent
northbound interface (NBI), intent policy layer, intent assurance layer and infrastructure layer, as
shown in Figure 5.

Figure 5 Intent-Driven Integrated Optical-Wireless Network Architecture

The service application layer generates service intent, including different services in different
scenarios. Service intent can be generated directly or indirectly. The service application layer
programs the underlying devices through the programming interface provided by the intent policy
layer, thus abstracting the functions of network elements. In addition, this layer provides
management interfaces to achieve diverse service innovation.
The intent northbound interface, a module used to translate intent, connects the service
application layer and the intent policy layer. In addition, the southbound interface (SBI) is based
on the virtualization technology and connected to various network element devices, virtualizing
various computing and communication resources. It is mainly used for interaction between the
18 / 56
infrastructure layer and the intent policy layer.
The intent policy layer is the core of the architecture with management control and
decision-making capabilities. This layer parses and checks the service intent translated through the
intent northbound interface. User intent is processed into a standardized intent request that can be
executed by the current network, and the specific resource requirements in the network are
obtained through a mapping algorithm between the intent and resources. This layer uses an
intent-based management and orchestration system for unified scheduling of resources and slice
configuration, and introduces the closed-loop configuration into the life cycle management of
network elements. With the intelligent engine, it completes network state data collection, data
storage, data processing, model training, parameter adjustment and other tasks. The infrastructure
layer includes various physical device entities. It also deploys a large number of network data
collection tools to provide feedback information and parameters required for policy configuration.
Optimized Design:
(1) Intelligent policy generation based on strong intent constraints (PG-RL)
By studying the extraction of intent keywords and designing intent request messages to
encapsulate the intent keywords, the intent can be expressed accurately. Under the constraints of
intent request messages, it studies policy generation based on reinforcement learning (RL), and, by
combining granular policies to generate new configuration methods for optical networks, optimal
adaptation of intent to policy can be achieved.

Figure 6 Architecture of the Knowledge Graph Matching System for Intent Translation
(2) Intelligent slicing policy generation based on intent constraints (SPG-RL)
By studying the mapping of intent and network resource requirements, it captures the service
intent and translate it into network policies, achieving accurate translation of intent into network
state requirements. Under the constraints of intent, it studies slicing policy generation based on
reinforcement learning (DRL), and, by effectively translating user performance requirements into
slice configuration policies for the integrated optical-wireless network, multidimensional sensing
requirements and optimal adaptation of intent to policy can be achieved, thus overcoming the
shortcomings of traditional optical network resource allocation policies—being static and
19 / 56
solidified.

Figure 7 Intelligent Slicing Workflow of the Intent-Driven Integrated Network

Figure 8 Intent Parsing Execution and Slicing Policy Result

4.4.2 Operation and Maintenance Methods and Policies

Deploy the network, and set the algorithms and computing nodes based on the intent-driven
optical network platform (IDONP) architecture. With intent assurance in mind, the maintenance
part studies the high-precision fault location algorithm and fast slice reconstruction algorithm for
intent assurance:
(1) To study the intent assurance mechanism based on high-precision fault location methods:
by introducing the deep neural evolution network, a high-precision fault location method for
large-scale alarm sets is proposed, which effectively enables the precise location of fault nodes in
optical networks, and helps the policy layer isolate faults in time to best assurance intent. The
simulation results show that if there are more than 10,000 alarms, the location accuracy is
increased to 92% and the location time is reduced by 0.5 seconds thanks to the embedded

20 / 56
deployment of the federated deep neural evolution network (FL-DNEN).

Figure 9 Improved Fault Propagation Model Algorithm

(2) The fast slice reconstruction algorithm for intent guarantee (SR-DDQN) studies the intent
assurance mechanism based on the fast slice reconstruction method. By introducing the Dueling
DQN network, a fast slice reconstruction method for highly dynamic network environments is
proposed, which effectively adjusts the policy of allocation of computing and storage resources in
different domains and the deployment of VNF to match the dynamic changes in slicing requests,
and thus to best assurance intent. The proposed model can continuously detect the consistency
between the execution policy and the original intent based on the real-time network state, and
adopt fast slice reconstruction when the intent requirements are not met. The reconstruction time
is reduced by 39.2%, and the resource utilization is increased by 17.6%, lowering the intent
constraint violation rate while improving network resource utilization.
By combining the above two algorithms and the structure of the existing network, the
network can realize automatic operation and maintenance. In addition, both algorithms have the
continuous learning capability. With the deployment and operation of the network, this
autonomous learning ability plus human supervision and regulation can help further optimize the
network.

4.4.3 Integration Methods Compatible with the Existing Network

This section briefly introduces the existing network structure, and how to integrate the IDN
and develop IDONP based on the existing network, describes the structure of IDONP and the
development process of each layer, as shown in Figure 10, and clarifies the workflow of IDONP.
All this offers integration ideas for IDNs.
(1) Introduction to the existing network: The software-defined optical network (SDON) is a
new network paradigm that enables flexible and efficient network management. SDON
technology allows the underlying infrastructure to be abstracted and used as a virtual entity by
applications and network services. This allows network operators to define and operate the logical
mapping of the network, and create multiple coexisting network slices (virtual networks) that are
independent of the underlying transmission technology and network protocols. SDON provides a
unified control plane platform to realize the integration of the access network, optical network,
metropolitan area network and core network segments, as well as the integration within and
between data centers. The global network view can be used to make optimal network control
decisions. SDON offers four main features, such as programmability, agility, flexibility and

21 / 56
independence to the network management domain. The idea of network functions virtualization
(NFV) is to virtualize network resources on shared facilities into virtualized network functions
(VNF). NFV allocates the capabilities of virtual machines (VM) or containers through a common
shared infrastructure with a VM manager, and directly controls hardware resources. The
SDON/NFV joint framework is the best way to realize intelligent control and orchestration of
intent-driven integrated optical-wireless networks. The SDN-based open network operating
system (ONOS) includes the intent framework component, which is designed to provide the
overall runtime environment and framework. As a subsystem of ONOS, the Intent Framework is
an integral part of intent connection. The Intent Framework views the service intent as a
policy-based directive, allowing applications to broadcast their network requirements externally
based on policy and management.
(2) Combined with the integration methods of the existing network: Based on the intent and
requirements, design the network architecture, including network topology, device layout, and
connectivity. Adopt a layered structure and modular design to facilitate future expansion and
optimization. The development of IDONP has the following practical significance:
Providing users with an interactive tool corresponding to the interaction layer in IDONP.
Users can issue intent, manage policies in the network, and obtain the network operating
conditions in real time through rendered web interfaces, APIs, or command line tools.
Providing a data model corresponding to the intent northbound interface (Intent NBI) in
IDONP. This model describes the detailed semantics behind the natural language and can
standardize the expression of semantics. Moreover, this data model is hierarchical, and a
high-level intent can be decomposed into several low-level intents.
Providing a policy engine corresponding to the JBoss Drools policy engine library in IDONP.
The policy engine separates granular network policies from applications, and can manage these
policies, including adding, deleting, modifying, and persisting them. This facilitates the policy
generation model based on reinforcement learning (PG-RL) to extract granular policies and
combine them to generate new network policies. IDONP can translate user intent into deployable
optical network policies to provide customized services that match user intent. And it integrates an
intent assurance mechanism based on the fault location method to form closed-loop control of the
intent life cycle, mainly including the interaction layer, intent layer, policy layer and fault location
layer.

22 / 56
 Figure 10 Design of the IDONP Structure
The interaction layer performs rendering mainly based on the HTML5 Canvas. It supports
infinite scaling of vector graphics, clearly and smoothly displays the operating conditions of the
optical network topology and online visualizes the returned back-end network data. The
interaction layer and the intent layer can, based on the HTTP protocol, issue the user intent to the
intent layer through Post requests, and obtain the intent translation result, policy generation
situation, and real-time network operating conditions through Get requests. In addition, the
interaction layer can also manage the policies in the JBoss Drools engine library of the policy
layer through APIs.
The intent layer receives the intent request issued by the interaction layer, parses the intent
semantics through the IA-LDA module, and obtains the intent keyword information. For example,
in the intent "It is necessary to ensure that sports events can be watched smoothly on the clients
during the World Cup", "QoE, to improve business priority, and to increase bandwidth allocation"
and other keyword information can be extracted. The intent NBI then encapsulates the keywords
into a json data structure containing the service information and issues it to the policy layer.
The policy layer includes the JBoss Drools policy engine library and the policy generation
model based on reinforcement learning. The policy engine library stores several granular rules,
mainly including, in IDONP, service sorting rules, routing rules, and optical network resource
allocation rules. Under the constraints of the data model issued by the intent NBI, PG-RL
combines the granular policies in the policy engine library to generate a new policy that matches
the intent.
In the fault location layer, FL-DNEN includes the fault propagation model and the DNEN
model. By monitoring the network and receiving alarms reported by the controller, it can
accurately locate the fault in the optical network and report it to the policy layer in a timely
manner, which helps the policy layer adjust service routing in a timely manner to avoid faulty
23 / 56
links or nodes and ensure effective intent assurance.
With all the above three aspects in mind, IDAN can be better designed, deployed, and
operated to offer more efficient, intelligent, and flexible network services.

V. Key Technologies Involved in the Full-lifecycle of IDAN

The core goal of IDAN is to translate the structured intent expressed by users into a set of
granular policies that can be executed by the network. Different from the traditional framework
that only views natural language as the source of intent, this technical white paper expands the
definition of intent to include various expression forms, such as action and voice, to capture the
user intent from more dimensions.

Figure 11 Intent Translation Process

Named-entity recognition, speech recognition, sequence annotation and other technologies
are used to accurately identify and extract key entities in the intent expression. Then, with
technologies such as text classification, semantic representation, and text generation, these entities
are mapped to predefined intent templates, thereby enriching and improving the intent expression
in the knowledge base. Afterwards, by using the policy library embedded in the knowledge base,
the enriched intent templates are translated into a set of specific sub-policies to form the policy set
required for execution.
Ultimately, the control layer will execute each sub-policy in this policy set in order, so as to
accurately implement the user intent. The overall process and structure of the intent translation
technology can be understood and implemented with reference to the technical flow chart shown

24 / 56
in Figure 11 to ensure efficient and accurate translation of user intent. To sum up, this chapter will
provide a detailed introduction to the key technologies involved in the above-mentioned spaces
and loops of IDNs.

5.1 Top-Down Intent Fulfillment

5.1.1 Intent Modeling

To achieve accurate mapping from user input to underlying network configuration, intent
needs to be classified. Human intent should not be classified from a certain perspective, as it is
multi-dimensional with different dimensions overlapping with each other. We can think of
describing an intent from multi-dimensions as "tagging" a certain intent. If there are sufficient tags
to fully cover the dimensions, the intent can be clearly described.
In terms of information modeling, the intent can be composed of a set of expectations, and
such expectations can be based on a common model [3]. When specifically defining the
expectation state list, the intent can be described based on the domain information model. For each
intent, there may be multiple dimensions of intent expectations and corresponding contexts, and
for each expectation, it may be composed of several expectation goals and a set of corresponding
expectation contexts.
In terms of data modeling, cross-system intent expression needs to use a
machine-recognizable language without ambiguous words for expression and exchange. For
example, the unified modeling language (UML) is a standard modeling language for software
development. It allows developers to describe, build, visualize and document the organizational
structure, behavior and interaction of software systems. The resource description framework (RDF)
is essentially a data model, and formally expressed as the subject-predicate-object (SPO) triple. It
is sometimes also called statements. YANG is a data modeling language with a tree structure
consisting of countless leaves, lists, leaf-lists, and containers.

Figure 12 Intent Classification Model

According to user goals and intent application scenarios, intent can be classified into
customer service intent, network service intent, network intent, operational task intent, policy
25 / 56
intent, etc. According to whether the name of the target object is explicitly specified, intent can be
classified into explicit intent and implicit intent. If the target component is clear, it is explicit
intent; otherwise, it is implicit intent. According to whether it has life cycle management
capabilities, intent can be classified into transient (operational) intent and persistent (service)
intent. Transient intent is a simple abstraction of network management operations. Once the
specified operation is executed, the intent is completed, and will no longer affect the target object,
so no life cycle management is required. On the contrary, for persistent intent, life cycle
management is required. Once the intent is activated and deployed, the system will keep all related
intent active until they are deactivated or removed.
In addition, IRTF RFC 9316 [5] proposed for the first time a standard model for intent
classification, which implements policies based on service scenarios, such as bearer networks,
cloud data center networks, and enterprise private networks, as well as various types of users, such
as network administrators and end customers, and can continuously and iteratively expand the
scale of the classification model as technology and services evolve. On the one hand, it improves
the intent recognition efficiency of the system. On the other hand, it provides differentiated
solutions based on user needs, expectations and priorities.
In order to describe it more clearly, intent can also be classified in terms of time, space,
whether it is periodic, etc. Figure 12 shows the model for multi-dimensional description of intent.
In terms of the level of intent, intent can be classified into basic intent and advanced intent. In
terms of time, intent can be classified into past intent, current intent, and possible future intent
(intent prediction). In terms of space, complex large-scope intent can be refined into small-scope
intent and refined from a global network configuration to a local network configuration. Intent can
also be classified into periodic intent and non-periodic intent. In terms of user service, intent can
be classified into the following types: information query, inter-node communication, network
configuration, and device control. In terms of urgency and importance, intent can be classified into
the following types: urgent and important, urgent but not important, important but not urgent,
neither urgent nor important, etc.

5.1.2 Intent Understanding

In IDNs, intent understanding is a key component that aims to translate human natural
language intents into structured intents. The basic process of intent understanding typically
involves the following steps:
1. User intent input: Users express their service intents through natural language (text or
voice) or graphical interfaces. Intents can range from simple commands to complex business
policy expressions.
2. Intent analysis: The natural language input is processed to extract keywords and phrases to
understand the specific requirements of user intents. This involves decoding the structure and
meaning of the input through syntax analysis and semantic understanding.
3. Semantic understanding and mapping: The extracted user language intents are mapped to a
set of predefined network operations and policies. Domain knowledge bases are used to link user
intents with specific operations that the network system can perform.
4. Intent verification and correction: The system checks whether the extracted intents align
with network policies and business logic. If necessary, the system will ask users to clarify or

26 / 56
correct their intent input.
Intent understanding is a critical part of an IDN because it is directly related to whether the
network can be correctly configured and managed according to the user's business goals and
requirements. Therefore, in order to improve the accuracy of intent understanding and the
adaptability of the network, it is usually necessary to combine advanced NLP technology, artificial
intelligence, and professional knowledge in the network field.
(1) Extraction model: GlobalPointer model construction
GlobalPointer is a span-based decoding method that considers the head and tail of an entity
as a whole, giving it a more "global point". It also ensures that training, prediction, and online
evaluation are all performed at the entity level, as shown in Figure 13.

Figure 13 Intent Entity Extraction Model

(2) Dataset construction: generation of over 1,000 datasets
(3) Entity categories: Entities can be transformed into intent models defined by various
standards organizations. Ultimately, these intents will be mapped to a set of predefined network
operations and policies. The information after mapping can be described as the following entities:
Objects (constraints on objects to which links apply): nodes, locations, regions
Behavior (constraints on how links are configured): connection, networking, firewall,
filtering, balance
Effect (constraints on link configuration effect): bandwidth, latency, throughput, packet loss
rate, QoS, business, network characteristics, protocol type
Time (constraints on link action time): action time
Additional information (additional information on link configurations): port, priority,
application scenario
Network
Network
Serial Operation
Operation Entity Example
Number Entity
Sub-category
Category
A1 Object Node Host1, A1, host, node, User B
A2 Object Location Point A, Point B
Subnet 1, ad hoc network, UAV network, satellite
A3 Object Region
network
Establish connections, enable connections,
B1 Behavior Connection
interconnect, communicate
B2 Behavior Networking Build subnets, form LANs
27 / 56
B3 Behavior Firewall Enable firewalls, disable firewalls
B4 Behavior Filtering Filter blacklists
B5 Behavior Balance Enable load balancing, enable controller balancing
Bandwidth reaches 500 Mbps, bandwidth not less
C1 Effect Bandwidth
than 500 Mbps
C2 Effect Latency Latency of no more than 5 ms
C3 Effect Throughput Maximum network throughput
C4 Effect Packet loss rate Packet loss rate less than 5%
C5 Effect Qos Highest QoS guarantee, QoS enabled
C6 Effect Business HD video conference, voice connection
Network Self-organizing, self-recovery, highly
C7 Effect
characteristics destruction-resistant
C8 Effect Protocol type IP mode, HTTP, SMTP
D1 Date Action time 8 a.m. to 8 p.m. every day for one week
Additional
E1 Port Port 439
information
Additional
E2 Priority Highest priority
information
Smart cities, wireless image transmission, interim
Additional Application
E3 conference, environment monitoring, emergency and
information scenarios
disaster relief

5.1.3 Policy Verification

Policy verification is responsible for ensuring the accuracy and applicability of policy
translation and generation to avoid potential risks such as network failures or service quality
degradation resulting from its direct implementation. This process requires policy verification to
not only be efficient in evaluating and filtering policies within the required timeframe but also
maintain a high level of accuracy to ensure that the selected policy can achieve the expected
results in practical applications. When verifying policy enforceability, we mainly focus on key
factors such as resource availability, potential policy conflicts, and the correctness of the policy
itself. At present, mainstream policy verification technologies typically involve two steps: building
a digital twin and conducting policy simulation and validation. These methods together form a
technical support system to ensure the reliability and effectiveness of the policy.
(1) Building a digital twin
Building a digital twin involves accurately simulating the physical network, which is a
process facilitated by advanced digital twin technology. This process is divided into two parts:
basic model construction and functional model construction. It aims to achieve a comprehensive
and dynamic description of the physical network. Basic models are created based on key data such
as the basic configuration of network elements, environmental information, operating status, and
link topology. This allows the creation of network element and topology models corresponding to
the physical entity network to ensure a real-time and accurate reflection of the physical network
environment. This process is further refined into three key steps: The first step is to build the
ontology model of the twin network to establish a unified and comprehensive network twin
28 / 56
database. Then, build the network element and topology models on demand to achieve an accurate
mapping of the physical network.
On the other hand, the construction of functional models focuses on meeting the requirements
for actual network functions. It supports the dynamic evolution and inferential decision-making of
the network by introducing diverse functional modules throughout the entire lifecycle. These
functional models can achieve comprehensive simulation and prediction of network functions
through multi-dimensional construction and expansion according to the specific requirements of
different network applications. The application of this technology not only enhances the flexibility
and responsiveness of network management, but also provides powerful support for network
planning and optimization. Figure 14 illustrates a schematic diagram of the digital twin of
Function C in the example.

Figure 14 Schematic Diagram of Digital Twin in Policy Verification

(2) Conducting policy simulation and verification
In order to improve the efficiency of policy verification and ensure the timeliness of policies,
a critical step is placed in front of the policy verification process: the pre-classification and
organization of the policies to be validated. This process involves using knowledge induction
methods to group the network policies based on shared characteristics (such as network protocols,
initial port numbers, and priorities), aiming to reduce the workload during the verification process
and accelerate the overall verification progress.
As the digital twin is successfully built, the network policies to be validated will be injected
into the twin for observation and analysis of simulation results. Policy verification typically
employs two main methods: control plane verification and data plane verification. These methods
respectively simulate the routing protocols at the network control layer and the data transmission
environment of the network to verify the reachability of routing policies and the efficiency of
network packet transmission.
Specific verification approaches can be further categorized into two types: global simulation
and local formal verification. Global simulation involves simulating the operations of all protocols
29 / 56
within a wide area network (WAN) to generate corresponding data plane processing rules. On the
other hand, local formal verification comes into play when uncertainties are encountered. It
encodes all potential scenarios with logical expressions and uses solvers to calculate various
possible situations. This allows for an accurate assessment of the effectiveness and potential
impacts of policies. This method provides a comprehensive and flexible technical path for
validating network policies, ensuring that the implementation of network policies is both efficient
and reliable.

5.2 Intent Realization Knowledge Management

Intent realization knowledge is used to standardize and associate user intent storage and
interaction update, intent translation result parameter storage, retrieval, and update, as well as
network state feedback information storage. Based on this knowledge, it is possible to achieve
closed-loop management of user intent and network state monitoring, providing feedback on the
closed-loop sensing of user intent. Meanwhile, intent realization knowledge management provides
a function for aggregating and exporting masking user information such as interaction information
on original user intent, intent translation result information, and network state feedback
information. It also establishes standard data specifications, which can serve as a foundational
dataset for the training and application of intelligent algorithm models related to intent networks in
the future. For the intent realization knowledge management model, an available resource library
needs to be designed to store intent realization knowledge and related information configurations,
and to handle intent translation results from the user interaction interface. A data collection and
analysis module needs be designed to query user intent processing results in the available resource
library and manage potential intent updates. Finally, a service orchestration module needs to be
designed to manage intent use cases, scenarios, and resources.
Intent realization knowledge management provides CRUD of intent instances. During the
creation of an intent instance, the user interaction interface reads the use case service ID, binds the
user intent to the newly created intent instance ID and the user service ID, and stores them in the
available resource library. Meanwhile, the available resource library offers a user intent
monitoring interface for the data collection and analysis module to listen for possible updates to
user intents. During the intent modification and update process, the user interaction interface reads
the new use case service ID, stores the new use case service ID for user intents in the available
resource library, and associates the ID with the existing intent instance. Intent realization
knowledge management also provides intent query and intent deletion functions.
By using intent realization knowledge management technology and mapping intent instances
to data tables (such as using an intent instance ID as the key ID), it becomes possible to associate
intent-related information. This allows for: (1) association of use case service IDs with intent
instance IDs in different scenarios to meet users' requirements across multiple scenarios; (2)
association of continuously changing user intents and intent translation information with a unified
intent instance; (3) association of network change information and intent assurance information
with a unified intent instance to assess and optimize the network's ability to support user intents.
Intent realization knowledge management technology also offers a function for aggregating
and exporting masking data, enabling the export of network intent-related information as a
foundational dataset for the training and application of intelligent algorithm models related to

30 / 56
intent networks in the future.

31 / 56
 VI. Application Cases of IDAN
The vision of intent-driven autonomous networks has become a consensus within the industry.
Various sectors such as telecommunications operators, equipment manufacturers, and higher
education institutions are actively promoting the practical application of intelligent networks. The
process of technology research and development and commercial application is accelerating, and
the entire industry is showing a trend of prosperity and development.

6.1 Application Cases - Intent-Driven CLL

China Telecom adopts the service-oriented concept of ONAP, focusing on 6G network
architecture design to address on-demand service requirements in all scenarios. China Telecom has
proposed and implemented an intent-driven closed-loop autonomous network architecture solution
and test platform and has validated key scenarios and technologies. In order to provide users with
differentiated and personalized intelligent services, ensure flexible scheduling of global cloud
network computing resources, and enable on-demand services in all business scenarios targeting
various intelligent use cases, China Telecom has implemented end-to-end integrated dynamic
orchestration management for 6G access networks, bearer networks, core networks, space-based
and air-based network domains, and cloud computing resource domains. In this example, user
intents from the service management layer are mapped to network execution policies at the
resource management layer, realizing intent-driven cloud leased line delivery and assurance.

Figure 15 On-demand Services in All Business Scenarios for Intent-driven Closed-loop

Autonomous Networks
From the perspective of closed-loop autonomous operations, the solution has designed a dual
closed-loop autonomous network architecture, as shown in Figure 15. The outer closed loop
includes two types of data flows, which respectively implement the functions of creating and

32 / 56
issuing intents and the functions of modifying and fulfilling intents. The inner closed loop
structure guarantees user intents and consists of four stages: M2A (monitoring), A2D (analysis),
D2E (decision), and E2M (execution). The tasks of each stage are defined and implemented by
modules in the solution architecture (such as the intent realization module, policy module, and
service orchestration module), and are specifically deployed and implemented in the ONAP
platform. This solution uses a lightweight decoupling method to achieve module decoupling, and
adopts standardized network intent verification processes and interfaces, demonstrating high
feasibility for implementation.
The intent input and translation module in the intent autonomous network is implemented
based on UUI. This module is responsible for parsing natural language input from users and
extracting corresponding network intent requirement information. The natural language input from
users can include voice-based or text-based network intent requirement information. Network
intent requirement information includes network quality of service requirements and basic
network configuration information. Network service quality requirements involve network
configuration parameters obtained through user intent awareness. Basic network configuration
information includes user location information. Network state information covers real-time traffic,
latency, and jitter. The voice-based or text-based network intent requirements from users can
undergo natural language processing methods (such as the BERT algorithm) for tasks like entity
extraction and recognition. Upon natural language processing, user requirements expressed in
natural language are translated into corresponding network parameters. SDNC and SO are
responsible for controlling SDN, fulfilling the intent, and then providing feedback to DCAE.
These policies are verified through the intent verification module to determine whether the
network state complies with the service level agreement (SLA). Finally, DCAE sends the
determination result to UUI for decision-making by users.
The intent instance management technology in the intent network enables closed-loop
management of user intents and monitoring of network state. When DCAE detects a change in the
network state, it sends a notification to UUI, requesting user intervention. Users can then modify
the intent instance parameters stored in AAI through UUI. Subsequently, AAI sends a
modification notification to DCAE to update the storage of user intents and intent translation
result parameters and read and update the storage of network state feedback information.
The Natural Language Processing (NLP) functionality in UUI categorizes user intents into
matching use cases or services, extracts parameters from the text, and populates them into
automatically generated service request forms. These forms are then presented to users for
confirmation or modification. A service form is a standard formatted expression of user intents,
which specifies the service type, requirements and SLA.
The intent orchestration management component in UUI reads the user intent parameters and
translates them into the corresponding IETF customer service model, which is then sent to SO.
The customer service model is a declarative model that describes the services provided by a
network operator to customers.
SO translates the customer service model into the corresponding IETF service delivery model
in the transmission network. The service delivery model defines how services are designed in the
network. SDNC converts the service delivery model into the corresponding network configuration
model and applies it to the physical network.
33 / 56
 Such model-driven design has its advantages. First, the model used in the translation step can
be standardized through SDO, which helps standardize IBN solutions. Second, the model
instances in each step are stored in the AAI database and can be retrieved through a RESTful
interface. Additionally, model-driven design decouples the data model from the code logic for
accessing (reading/writing) data. This decoupling enables us to improve and evolve our code logic
independent of the data model, thereby minimizing impacts on other components.
The inner closed loop of intent guarantee consists of 4 stages: monitoring, analysis,
decision-making, and execution. The monitoring stage is implemented on SDNC. It collects
monitoring and performance data from the network controller and forwards the data to DCAE.
SDNC determines which data to collect based on SLA parameters obtained from user intents
during the intent translation process.
In the analysis stage, DCAE analyzes the monitoring data received from SDNC and provides
feedback to policy. Its significance is to detect network anomalies and notify policy to take
corrective measures.
The decision-making stage is implemented on policy. It makes closed-loop decisions based
on the data received from DCAE and issues appropriate recommendations to SO to perform
service changes.
The execution stage involves a typical SDN orchestration and control workflow. SO and
SDNC apply new network configurations to the physical network.
All data, including monitoring data, business models, and network resources and
configurations, are stored in AAI for retrieval and sharing at various stages.
The Cloud Leased Line (CLL) service connects cloud service users to edge clouds or cloud
data centers and connects edge nodes to each other using connection technology or
connection-oriented technology to provide deterministic connectivity performance. Figure 16
shows an example of CLL service architecture using Ethernet-over-OTN (EOO) technology. User
data (such as IP network packets or Ethernet frames) is transmitted through bearer Ethernet
services, which run on Ethernet Virtual Connection (EVC). EVC provides the operations,
administration, and maintenance (OAM) of the services, while the underlying optical transport
network (OTN) service provides service isolation and traffic protection.

Figure 16 EOO-Based CLL Service Architecture

Figure 17 shows the actual deployment of the IDN architecture . The physical network is a
dual-domain optical network, with each network domain controlled by a physical network
controller (PNC). The standard interface between ONAP and PNC complies with the IETF/ACTN
standard (multi-domain service model and protocol interface).
34 / 56
 Figure 17 Cloud Leased Line Deployment Diagram
Figure 18 illustrates the model-driven intent translation process. First, the user's intent
expressed in natural language (as shown in Figure 18(a)) is translated into an intent model (Figure
18(b)) through the NLP function. NLP translates the user-expressed endpoints "campus A" and
"cloud 2" into the corresponding endpoint names in the network topology
(transportEp_src_campus_A1 and transportEp_dst_cloud2) for SO and SDNC to comprehend.
Next, the intent orchestrator in UUI translates the intent definition into a customer service
model, which is a specification of the CLL service the user wants to create (Figure 18(c)). The
customer service model specifies the connectivity and Service Level Agreement of the CLL
service in a technology-agnostic manner. Finally, the customer service model is translated into a
network configuration model, as shown in Figure 18(d). These models represent the actual
network configurations of EVC and its underlying OTN tunnels required to implement the CLL
service.

Figure 18 Model-Driven Intent Translation

35 / 56
 Once the intent is fulfilled, UUI presents the configured network service to the user, as shown
in Figure 19. It assists the user in inspecting and validating whether the network aligns with their
intent. In addition, the specific content of intent guarantee can also be displayed in UUI. Figure 20
shows the closed-loop process in a CLL use case. It can be seen that the actual bandwidth usage
can be retrieved by the system. When the actual bandwidth exceeds the preset bandwidth, the
system can proactively adjust the maximum bandwidth to avoid traffic congestion.

Figure 19 CLL Service in UUI

Figure 20 Closed-Loop Guarantee

The blue curve represents the actual bandwidth usage monitored by the intent-driven
autonomous network solution, and the green curve represents the pre-configured maximum
bandwidth. When a bandwidth alarm is triggered, the system can send a notification to UUI to
request user intervention. Users can choose to modify the original intent to increase the maximum
bandwidth.

6.2 Application Cases - China Unicom

6.2.1 Intent-driven Autonomous Orchestration Case

China Unicom has joined forces with manufacturers and provincial branches to adopt the
concept of "user experience as the center" to create the industry's intent-driven 5G endogenous
intelligent orchestration network applications, and innovatively implement the "user
experience-centered" intent-driven approach based on network endogenous intelligence. Practical
applications of the network. This solution is driven by user intents and relies on 5G-A-oriented
endogenous intelligence. Through intelligent means, it achieves dynamic network policies and
resource configuration based on business requirements to meet the evolving demands of diverse
network services in the future [12]. In this example, the extrinsic intent from the service
management layer is mapped to network execution policies at the resource management layer,
enabling intent-driven autonomous orchestration network applications.

36 / 56
 Figure 21 Intent-Driven Autonomous Network Orchestration
Compared with traditional solutions, this solution boasts characteristics such as intent-driven
operation, experience-driven design, endogenous intelligence, precise orchestration, and agile
switching. The solution can significantly improve user experience for 5G services such as
high-definition live streaming, video calls, video ringback tones, and AR/VR. In the past, in areas
with poor 5G coverage and high interference, these services were prone to frequent screen blur
and freeze, leading to a drop in user perception. Intent-driven intelligent orchestration ensures that
network resources can always meet the service needs of users and greatly improve user
perception.
The intent-driven autonomous network orchestration solution leverages the extrinsic
intent-driven knowledge base within 5G base stations, comprising both a business knowledge base
and a logical raster knowledge base, as shown in Figure 21. The role of the business knowledge
base is to assist in accurately identifying user business requirements and making precise
judgments regarding the user experience in the source cell. Meanwhile, the logical raster
knowledge base aids in anticipating and precisely selecting the experience in the target cell,
facilitating seamless and agile switching. Through the precise identification of user business
requirements, precise judgment of the user experience in the source cell, anticipation and precise
selection of the experience in the target cell, and seamless and agile switching, this solution
implements intent-driven flexible orchestration services, thereby ensuring differentiated user
experiences.

6.2.2 Intent-Driven Autonomous Energy Saving

As an important foundation for the development of the digital economy, the 5G network has
become a key driver in the national "dual carbon" strategy. Promoting the green development of
5G is a fundamental requirement for ensuring the sustainable development of resources and the
environment and is of great significance to the comprehensive green transformation of the
economy and society. In order to implement the national "dual carbon" strategy and meet the

37 / 56
requirements for green 5G development, China Unicom has created a practical case of IDNs with
"green and low-carbon" as its core.

Figure 22 Intent-Driven Network Autonomous Energy-Saving Solution

In traditional energy-saving solutions, it is difficult to achieve intelligent energy-saving
across multiple systems and maximize the overall network energy efficiency, which is an
important challenge in building green 5G. Compared with traditional solutions, this solution is
driven by the intent of autonomous network energy saving. It uses autonomous means to achieve
intelligent energy saving in RANs in multiple modes and at multiple sites, meeting the
requirements for intelligent, refined, and large-scale closed-loop management of network
energy-saving. In this example, the energy saving intent from the resource management layer is
mapped to energy saving execution policies at the network element management layer, enabling
intent-driven autonomous energy saving.
Driven by the intent of network energy-saving, the IDN autonomous energy-saving solution
relies on various network energy-saving scenario models, traffic prediction models, and network
deployment methods. It combines the preset operator policy intent tendencies to achieve the
automatic and precise selection of energy-saving policies and optimization of energy-saving
parameters. As shown in Figure 22, the solution forms a closed-loop process of intent awareness,
intelligent analysis, flexible decision-making, and automatic execution. This process aims to
reduce energy consumption while ensuring network performance.

6.2.3 Automatic Drive Test Intelligent Evaluation Solution

Traditional drive test mainly relies on manual intervention, which is time-consuming,
labor-intensive, and lacks intelligence. China Unicom's automatic drive test intelligent evaluation
solution makes full use of the capabilities of intent-driven autonomous networks. The integration
of "testing, evaluation, calculation, and management" revolutionizes the traditional drive test
mode and reduces the labor-intensive workload of drive testing and analysis. In this example, the
intent of the automatic driving test and intelligent evaluation from the service management layer is
mapped to execution policies at the resource management layer, enabling intent-driven

38 / 56
autonomous driving test.

Figure 23 Automatic Drive Test Intelligent Evaluation Solution Flow Chart

As shown in Figure 23, the solution is deeply empowered by IDNs. It automatically conducts
problem identification, root cause analysis, intelligent decision-making, automatic execution, and
closed-loop evaluation of execution effects, thereby enabling end-to-end management of
problematic road sections and achieving Level 3 autonomy level in this scenario:
(1) Problem sensing: The solution automatically collects massive network MR/MDT data,
uses AI models to fit key indicators of road conditions, and achieves automatic collection,
processing, and geographical presentation of four major categories of indicators, including
coverage, quality, voice sensing, and data sensing. Based on this, problematic road sections are
automatically identified.
(2) Root cause analysis: The solution conducts root cause analysis of identified problematic
road sections, quickly presents the analysis results of problematic road sections, and shortens the
analysis time from 2 hours required by traditional methods to 15 minutes, significantly improving
analysis efficiency.
(3) Intelligent decision-making: Based on the analysis results of problematic road sections,
the solution uses the decision-making AI model to provide optimal solution suggestions, including
specific optimization measures and parameter tuning suggestions.
(4) Automatic execution: The system can flexibly connect to the ticket system, directly push
root cause analysis results, handling suggestions, and key indicators to the front line, helping the
front line to quickly solve problems.
The system automatically evaluates the effectiveness of the optimization results after a ticket
is submitted, and judges whether the problem is solved, and how effective the solution is, thereby
achieving end-to-end closed-loop control of problematic road sections. At present, this scenario
has been applied on a large scale in 31 provinces, with a total of 5.21 million kilometers of 4G/5G
test-free mileage, equivalent to a test cost of CNY 109 million. After actual verification of existing
networks in many provinces, the matching accuracy of road quality problems found in traditional
manual road tests is over 90%.

6.3 Application Cases - China Mobile

6.3.1 Intent-driven Hierarchical Service Assurance Case

Current network service assurance faces several challenges: poor user experience due to the
high skill requirements of traditional operations, long response time caused by numerous O&M
parameters and complex processes, lengthy feedback cycles lacking timeliness, rigid scheduling
policies unable to flexibly meet diverse business needs, and the inability to provide continuous
assurance covering the entire lifecycle of services. To address these issues, meet the differentiated
39 / 56
business needs of users, and enhance the flexibility of business services, we have adopted an
intent-driven service assurance method. This method aims to promptly and efficiently address the
diverse needs of different users, thereby improving the overall business experience. In this
example, the intent for hierarchical service assurance from the business management layer is
mapped to differentiated scheduling policies at the service management layer, implementing
intent-driven differentiated service assurance.
Intent-driven hierarchical service assurance involves recognizing intents, matching them with
different levels of assurance capabilities, and providing differentiated guarantees for network
service quality. These include the following different hierarchical assurance capabilities:
(1) Gold/Level 1 assurance: strictly restricted usage. This level of assurance supports
unlimited wireless resource preemption, supports intelligent scheduling policy optimization, and
enables all high computing power overhead functions.
(2) Silver/Level 2 assurance: only targets key service assurance. This level of assurance
supports limited wireless resource preemption, supports intelligent scheduling policy optimization,
and enables high computing power overhead functions on demand.
(3) Copper/Level 3 assurance: on-demand usage. This level of assurance does not support
wireless resource preemption, but supports intelligent scheduling policy optimization, and enables
compute-intensive functions on demand.
Multi-dimensional differentiated services are provided for the preceding different hierarchical
assurance capabilities:
1) Differentiated scheduling priorities: Diverse priority levels are supported based on
different 5QI, NSSAI, and types of services. This includes facilitating wireless resource
preemption for higher-priority services. 2) Differentiated scheduling policies: Distinct scheduling
parameter configurations, such as configurations for the target BLER and uplink power control,
are supported based on different 5QI, NSSAI, and service types. 3) Differentiated scheduling
algorithms: High computing power overhead functions such as enhanced LDPC iterations and
PUCCH format1 IRC can be enabled for specific 5QI, NSSAI, and types of services.
In practical terms, the intent-driven hierarchical service assurance system follows the
following process:
(1) Intent input: The user inputs sensing service logic in natural language, including the
region, app, time, and assurance level.
(2) Intent translation: checks whether the sensing intent input is correct, guides the user to
input intent, and translates the user intent into computer instructions.
(3) Conflict detection: checks whether the sensing service can guarantee intent execution,
checks for conflict with other services, and provides modification suggestions.
(4) Intent execution: ensures that the user app is perceived with differentiated scheduling
policies based on specific 5QI, slice ID, and service and parameter policy orchestration.
(5) Intent protection: rapid assessment and continuous monitoring of the effect of sensing
intent assurance. Once user requirements cannot be met, an alarm is sent.
Providing intent assurance capabilities based on the preceding process can effectively
improve key metrics of the app, including opening latency, the success rate of sending/receiving
images/videos, and the effective video download rate. This further enhances the user's trust and
reliance on the app, thereby promoting the long-term development and user growth of the app.
40 / 56
Take the service assurance for a specific app as an example. After intent assurance capabilities are
implemented, the average video download rate is increased by 0.387 Mbps within one week,
representing a growth rate of 1.32%. Moreover, there is a noticeable improvement in the average
number of video playback freezes, with an improvement rate of 21.32%, as shown in Figure 24.

Figure 24 Comparison of the Average Number of Video Playback Freezes Before and After
Deployment of Intent Guarantee Capabilities
The practical effectiveness of intent-driven hierarchical service assurance is evident in its
meticulous categorization of services and the precise formulation of assurance policies based on
the service intent at each level. This approach not only ensures the continuity and stability of core
operations but also optimizes resource allocation, reflecting the following values:
Extreme user-friendliness: provides a friendly interaction method for controlling the network
through natural language. Users only need to communicate their network management objectives
and constraints without requiring knowledge of implementation methods.
Rapid response: minute-level response and assessment, millisecond-level service
identification and assurance policies, and more than 500% improvement in assurance efficiency.
Excellent performance: improved guaranteed service/UE performance in congestion
scenarios, up to 50% reduction in latency, and 100% throughput increase.
Ultimate assurance: flexible scheduling, and service assurance covering the entire life cycle.

6.3.2 Intent-driven Sensing Certainty Assurance Case

The intent-driven sensing certainty assurance function jointly introduced by Guangxi Mobile
and ZTE was successfully applied in the 5G network for the first time [13]. This function analyzes
user intent to accurately identify specified video services. It then dynamically monitors and
efficiently allocates network resources in real time to provide intelligent assurance for the
specified video services based on the target Mean Opinion Score (MOS) value, which represents
the subjective experience of video quality. In this example, the user intent from the business
management layer is mapped to resource allocation configuration policies at the resource
management layer, enabling intent-driven video service assurance.

41 / 56
 Figure 25 Intent-Driven Sensing Certainty Assurance
As shown in Figure 25, the initial intent of O&M personnel is analyzed by an intent engine
and transformed into network-executable policies and instructions. This process also supports
real-time monitoring and adjustment of policies. The intent-driven sensing certainty assurance
function uses artificial intelligence technologies such as deep learning to analyze a vast amount of
video service data and establish models for video classification and quality assessment. In the 5G
network environment, this function can dynamically perceive the category of the app and its
network resource requirements. Then, it efficiently allocates resources and selects paths for
specified video services through technologies such as intelligent scheduling algorithms and
priority techniques, thereby achieving differentiated quantitative assurance. This differentiated
assurance experience for specified services maximizes the technical advantages of the 5G network.
Figure 26 shows the specific process of the intent-driven sensing certainty assurance solution.

Figure 26 Process of Intent-driven Sensing Certainty Assurance

The economic benefits of the case are primarily reflected in two aspects: saving manpower
and resource costs for assurance and generating additional revenue from increased traffic at
assurance sites. In traditional approaches, assurance for a specific service often necessitates
collaborative efforts from a team. This involves various tasks such as conducting coverage
assessments, estimating capacity requirements, adding temporary sites (as needed), configuring
data settings, and ongoing monitoring of performance metrics, faults, and alarms during the
assurance process. On average, each day of assurance typically requires about 8 person-days of
work. With the introduction of this case, only one person is needed to issue assurance instructions,
reducing the manpower required for assurance by 7 person-days per day. Intent guarantee also
minimizes the impact of traditional guarantee policies on the network, avoiding negative impacts
and problems caused by improper operations. This, in turn, reduces the need for troubleshooting
manpower, further lowering carbon emissions.

42 / 56
 6.3.3 Autonomous Intent O&M Assurance High-quality Industry-specific
Leased Lines
The intent-based convenient service provisioning and self-maintenance solution is a
successful practice in the intelligent O&M of industry-specific leased lines. China Mobile of
Zhejiang Shaoxing, in collaboration with ZTE Corporation, successfully conducted intent
maintenance tests in service degradation scenarios. The test results were positive, showing
significant effects in cost efficiency and customer perception improvement. This lays a solid
foundation for comprehensively elevating the service level of leased line services across the SPN
network in the future.
In response to the demand for the rapid launch of leased line services and ultra-high quality
assurance, China Mobile of Zhejiang Shaoxing has proposed an intelligent intent O&M solution
for industry-specific leased lines based on the intelligent control system. This solution better
adapts to customer needs by facilitating convenient service provisioning and proactive
maintenance of service quality. It allows O&M personnel to focus solely on customers' SLA
intents for services. The intelligent control system automatically completes subsequent operations
based on the intents and verifies intent matching. In this example, SLA intents from the business
management layer are mapped to policies such as service provisioning and quality assurance at the
service management layer. This integration enables intent-driven intelligent O&M of
industry-specific leased lines.
The main application scenarios of this solution are as follows:
1. Intent-based convenient service provisioning: As shown in Figure 27, the function of
automatic provisioning of services based on user intent involves a fully automated and visualized
process that eliminates the need for the tedious and error-prone configuration of basic business
information. Users can simply articulate their specific requirements for leased line services
through the UI, voice commands, or text input. The system uses intent recognition to trigger a
guided input process, prompting users to specify details such as device endpoints, bandwidth, and
protection requirements. Subsequently, the system automatically generates multiple service
solutions that align with the user's intent. These solutions undergo simulation to assess network
service quality under various policies and scenarios. Finally, the system selects the optimal
recommended solution, and users confirm their choice from the proposed service solutions. The
system converts the solution content into various configuration details for the devices involved.
After undergoing necessary verification processes within the system, these configuration details
are then distributed to the relevant devices. This completes the provisioning of the requested
services.

43 / 56
 Figure 27 Intent-Based Convenient Service Provisioning and Self-maintenance Solution
2. Intent-based service self-maintenance: With the idea of the IDN as the core, the system
matches corresponding monitoring policies based on the SLAs for customers' leased line services.
It uses in-band OAM for in-situ flow information telemetry, enabling real-time perception and
monitoring of service quality. This includes monitoring quality metrics such as packet loss,
latency, and jitter for individual service flows. The collected telemetry data is reported to the
intelligent control system within seconds. The big data platform stores and correlates this data,
while the sensing engine conducts real-time analysis and evaluation. It verifies whether the service
quality aligns with customer intent in real time, and continuously adjusts to form a closed-loop
system for service quality control assurance.
3. Accurate sensing of service state and real-time maintenance of service quality: In-situ
flow information telemetry enables multi-dimensional visualization of service quality and precise
network insights, showcasing relevant service quality and traffic information of leased lines. This
helps users quickly understand the status of network service quality. By actively monitoring and
combining the overall network state and big data analysis and prediction, the system proactively
perceives potential faults before they occur. This allows for preemptive handling of potential
network issues, enabling rapid self-healing of services and guaranteeing service bearer quality.
4. Always-on services and quality restoration without any customer awareness: Accurate
sensing of SLA quality and three-dimensional network protection enable comprehensive service
assurance. Whether it is a single point or multiple points of failure, the controller and device layer
IGP collaborate to realize service rerouting and escape paths for multi-dimensional protection.
This mechanism allows for the recovery of services within 50 milliseconds, achieving
telecom-grade protection standards.
The solution was tested and verified at Zhejiang Shaoxing Mobile. Based on actual test
results, it has the following values:
1. Reduction in O&M costs: The end-to-end O&M of leased line services is streamlined to
reduce the involvement of O&M personnel, simplify service configuration, and ensure automatic
service quality assurance through a business quality system. This leads to savings in O&M
manpower investment. Through automatic real-time monitoring, dynamic maintenance, and
continuous optimization of service quality, the system ensures that services are permanently
online and maintained at a high quality. This enhances the level of automation in service
maintenance, reduces reliance on individual experience in business quality analysis, simplifies
44 / 56
O&M, and consequently reduces O&M labor costs. Taking the maintenance of leased line services
in a single city as an example, the average number of related maintenance events is about 90 per
month, with each event requiring approximately 6 hours to handle. This totals around 22
person-days of work per month. With the introduction of the service self-maintenance system, the
total working hours are reduced to 9 man-days per month, reducing maintenance working hours
by 60%.
2. Improve maintenance efficiency: The service state can be perceived in real time and
displayed graphically on the maintenance interface promptly to directly reflect the operation status
and change trend of the network and service. The self-maintenance of service quality and timely
feedback on the service state can avoid inaccurate identification of manual queries and improve
maintenance timeliness.
3. Improve user experience: The intent-based self-maintenance solution for leased line
services provides 24/7 automatic closed-loop quality assurance services throughout the service
lifecycle. It reduces the O&M load and fundamentally solves the problems of inadequate services
and unguaranteed service quality caused by passive O&M.

6.3.4 Implementation of General Intent Solutions via ONAP

In an autonomous network, it is necessary to define a set of general intent solutions for
intents between different domains/layers. Specially, a general intent information model and intent
interaction interfaces should be defined, so that intents in the network can be described using the
same model and follow a unified interaction process. Based on the global existing intent-related
standards, China Mobile has referenced the practices of intent-driven autonomous networks to
lead the project of general intent model and interfaces in the LFN ONAP community. By studying
the key technologies, implementations, and industrial standards of IDN, China Mobile guides the
community to implement general intent solutions. In the application case mentioned above, the
intent of the on-demand guarantee from the service management layer is converted into the scaling
policy at the resource management layer to realize the intent-driven high-quality service guarantee
of cloud leased lines.
In the specific business, the end-to-end intent realization process is demonstrated through the
cloud leased line scenario, covering the delivery of cloud leased line service and on-demand
automatic guarantee. Operators provide intent-based cloud leased line services and take
corresponding guarantee measures according to user needs. For example, an enterprise user
purchases a cloud leased line with 1 G bandwidth. To satisfy this intent, the intent management
function configures a cloud leased line with 1 G bandwidth according to the intent handling
process. Some advanced quality of service guarantee technologies are also provided. For example,
for high-level users, the bandwidth will be automatically expanded by 60% to ensure user
experience when the bandwidth utilization exceeds 80%. In addition, the bandwidth will return to
1 G when traffic becomes normal again (utilization of 30%).
In response to the above business needs, the solutions are as follows:
1. Intent analysis service:
In 4.1 Hierarchical Operation Architecture of Intent-Driven Autonomous Network, it is
proposed that an overlayed intent layer can be introduced at all layers of the autonomous network
to realize inter-layer intent interaction. Based on the existing architecture of ONAP, China Mobile

45 / 56
has developed the intent analysis service. As an intent management function at the business
management and service management layers, it provides ONAP with intent processing capabilities,
especially a unified intent handling process for intent interaction between different domains/layers,
as shown in Figure 28:

Figure 28 Architecture of ONAP with Intent Management Function

The introduction of the intent analysis service provides the following functions:
 Description and interpretation using the general intent model: The machine-machine intent
formatted between different domains/layers can be described and analyzed using the general
intent model.
 General intent interfaces: A unified intent interaction method and process are provided for
other or third-party components.
 Intent translation: Intents described using the general intent model can be translated into
specific execution policies or actions.
 Intent splitting and orchestration: Complex intents can be split into sub-intents for
orchestration.
 Formatted intent input interface: A unified and formatted intent interaction and input
interface is provided.

2. The general intent information model and intent interfaces:

At present, intent standard projects of organizations such as CCSA[3], TM Forum[4], and
3GPP[6] have proposed intent information models and intent interaction interfaces. The team of
China Mobile has analyzed the similarities and differences among these models and interfaces. On
this basis, the team has proposed a general intent model and interfaces that conform to the ONAP
architecture.
The team innovatively proposes the Condition model based on the existing information such
as intent expectation, expected object, expected objectives, and context. This model can give
46 / 56
flexible condition expressions for complex expected objectives and contexts, which helps realize
complex intents that conform to actual autonomous networks. For general interfaces, the team
implements the negotiation interface except for existing CRUD interfaces. The intent owner and
intent processor negotiate the modification of intent parameters to ensure the realizability of
intents.
3. Intent-driven end-to-end implementation process:
In specific implementations, the intent input is formatted to simulate the user's intent. The
corresponding information is then sent to the intent analysis service. The intent analysis service
splits the user's complex intent into two sub-intents: service delivery intent and service guarantee
intent. For the delivery intent, the intent analysis service calls the service orchestrator to create a
cloud leased line with specified specifications. For the high-level quality of service guarantee
intent, the intent analysis service calls the corresponding guarantee policy to perform bandwidth
scaling on demand, ensuring the user experience of high-level users. The specific process is as
follows:

Figure 29 End-to-End Implementation Process of Cloud Leased Line Service

As shown in Figure 29, the solution ensures that the high-level user experience can be
satisfied by realizing the service guarantee intent according to the actual network situation while
realizing the intent of activating the cloud leased line service from users. The solution can be used
to implement general intent management based on ONAP. The general intent model standards are
released in the community, and related research results are also promoted and applied in intent
standard projects of CCSA, ETSI, and other organizations, which enhances industry cooperation
and promotes open-source operations and collaboration of standards.

6.4 Application Cases - Emergency Communication Network

6.4.1 Case Analysis of Call on Session Initiation Protocol (SIP)

47 / 56
 The cross-domain communication between the self-organizing network and the cluster
network is shown in Figure 30. The user's mobile phone can connect to the self-organizing
network and make voice calls with the walkie-talkie in the cluster network. The user information
on the mobile phone and walkie-talkie is registered in the SIP server, and the walkie-talkie is
connected to the cluster gateway. The mobile phone is wirelessly connected to a self-organizing
network device with Wi-Fi turned on. The SIP server, router, cluster gateway, and self-organizing
network device support the Internet protocol. Therefore, they are physically connected. To make a
voice call, an appropriate route and session parameters should be configured at the network and
application layers, respectively. Assume that the IP address of the SIP server is 192.168.30.30,
which is in the destination network 192.168.30.0 of the router egress port eth0. The IP address of
the self-organizing network device is 192.168.168.41, which is not in the destination network
192.168.1.0 of the router egress port eth1. The IP address of the cluster gateway is
192.168.166.205, which is in the destination network 192.168.166.0 of the router egress port eth2.
If the intent driving proxy receives an intent of "call between the mobile phone and walkie-talkie",
it will parse this high-level intent into a low-level intent of "modifying the destination network of
the router egress port eth1 to 192.168.168.0". Then, the router configuration will be modified by
translating the intent and applying the policy. After the above operation is completed, the mobile
phone can communicate with the walkie-talkie through the LAN. This solution is a feasible
technical implementation solution, but the specific deployment needs to be carried out in
accordance with relevant management regulations.

Figure 30 Case of Call Request Intent

6.4.2 Integrated Space-Air-Ground Communication Network for

Emergency Rescue
In Figure 31, a unified architecture of integrated space-air-ground communication network
for emergency rescue is provided to meet the needs of emergency rescue communication in
response to forest fires, earthquakes, geological disasters, and other circumstances. The core
component of the network is the MEC gateway (intent-driven proxy), which is mainly responsible

48 / 56
for intent driving, protocol conversion, and mobile edge computing. In addition, the intent-driven
controller and MEC gateway are connected through a southbound interface to generate an intent
for cooperation in case of emergency. For example, when a forest fire occurs in a place, the
satellite link system quickly detects the specific area of the fire and transmits the corresponding
data to the MEC gateway. At this time, the intent-driven controller uses the received data to make
predictions and initiate emergency rescue communication between the cluster network and the
self-organizing network. This process involves real-time fire perception, intent prediction, and
communication network utilization, ensuring efficient and rapid rescue operations during an
emergency. The collaboration of the whole system enables the network to respond quickly and
provide required communication resource support in case of disasters. The mobile edge computing
capability of the MEC gateway allows it to process part of the data locally to reduce the load on
the central server and improve communication efficiency. The architecture of an integrated
space-air-ground communication network for emergency rescue provides reliable technical
support for emergency response in case of disasters.

Figure 31 Architecture of Integrated Space-Air-Ground Communication Network for Emergency

Rescue

VII. Technical and Application Challenges of IDAN

7.1 Technical Bottleneck Analysis

In the future, network development should focus on intelligence and be based on service
demand perception. Network and computing resources should be dynamically and uniformly
orchestrated to provide customized network services and diversified service expansion capabilities,
realize the deep integration of intent and intelligence, and offer end-to-end customized capabilities
to dynamically deploy and use network resources on demand for diversified business applications.
In this way, network development can achieve the goal of integrating network resources and
empowering emerging businesses.
49 / 56
 7.1.1 Uncertainty of Global Information
The intent-driven controller needs to know global or local network information to complete
the specific network configuration based on user intents. When the intent scope is wide, it is
difficult for the intent-driven controller to obtain global network information, such as the global
network topology information. This is an urgent problem for communication network studies. To
solve this problem, hierarchical management and maximum optimization can be adopted. When
the intent scope is wide, the lower-level intent-driven controller interacts with the higher-level
intent-driven controller for hierarchical network control to optimize the local network
configuration and state.

7.1.2 Variability of Dynamic Intent

User intents can be divided into instant intents, short-term intents, and long-term intents
based on the lifecycle. Changes along with the time are usually not considered for instant and
short-term intents. For long-term intents, as time goes by, although the business purpose remains
the same, the intent and underlying network configuration may change greatly due to network
state changes and the emergence of new technologies. Therefore, to cope with the dynamic
evolution of long-term intents, more advanced AI technology and dynamic network configuration
technology are required for processing users' intents. Figure 32 shows the evolution of intents.

Figure 32 Intent Evolution

7.1.3 Full-Lifecycle Closed Loop

Considering the complexity and variability of intents, this whitepaper puts forward the
concept of a full-lifecycle closed loop. In an intent-driven autonomous network, the team believes
that each intent has its own lifecycle. The end of the lifecycle means the end of the intent. The
system will clear related resources and mark the intent as completed or obsolete. Simple intents
have a short lifecycle, and changes are often not considered. However, for more complex intents,
great changes may occur. For example, the intent may be converted into another intent, optimize
or end itself autonomously, or be merged with other intents. The design of the intent lifecycle
helps the system better adapt to complex and dynamic environments and ensures that the system
can flexibly and intelligently deal with various tasks and situations.
In-depth research on the full-lifecycle closed loop of intents is still necessary in many aspects.
Researchers can explore the automatic detection and incubation of new intents, effective
evaluation and negotiation of feasible intent implementation solutions, precise intent delivery to
correct handlers and successful instantiation, continuous achievement of expected goals of intents
during implementation, and handling of intent update and deletion requests to solve possible
conflicts.
The research will face the main technical challenges of achieving efficient collaboration
50 / 56
between intent owners and handlers, ensuring the evaluation feasibility and effectiveness,
continuously achieving the expected goals during implementation, and ensuring that intent
processors can effectively report the completion and execution states.

7.2 Aspects to Explore in the Future

7.2.1 Bottom-Up Intent Assurance

IDN is an emerging network technology and methodology. Its full lifecycle can be divided
into user space, digital space, and physical space. The user space allows users to declare intents,
report network state, and evaluate network performance. The digital space bridges the user space
and physical space and contains key steps such as translation, optimization, analysis, and feedback.
It can verify intents based on intent demands so as to take corrective measures promptly. In the
physical space, the physical and virtual network elements are decoupled from each other by
functions and serve as independent functional modules. The full-lifecycle closed loop of intents is
shown in Figure 33.
The entire intent lifecycle contains three key loops. In Loop L1, intents are presented by
applying configuration and performing operations, and the network is monitored and analyzed.
During monitoring, information about policies and configurations is obtained from physical
devices in the network in real time for policy extraction. If the extracted policy does not comply
with the generated logical policy, feedback is given to adjust the network configuration. In Loop
L2, intent policy reasoning is carried out based on the content monitored in the physical space. If
the output does not match the policy for user intent translation, feedback is given to adjust the
logical policy generation. Otherwise, the output can continue execution directly as a "reasoned
intent". Loop L3 involves three spaces, covers the entire process from intent generation to intent
ending, and achieves intent realization and guarantee. It also includes the process in which users
take actions based on feedback from the digital space and adjust the intent input.
A bottom-up intent assurance ensures that the needs of intent owners are met
comprehensively and efficiently. This mechanism includes three core objects. The first one is the
internal closed loop of the intent processors. With rigorous logic and process, intents can be
delivered accurately. According to the actual situation of indicators corresponding to the intent
expectation, the intent realization state can be optimized and iterated based on policies and other
technologies to improve the efficiency and quality of intent realization. The second one is the
mechanism of continuous intent reporting. The intent processor regularly feeds back the intent
realization status to the intent owner using reports. These reports contain the progress and results
of intent realization and analyze existing problems and challenges in detail to provide
comprehensive information support for intent owners. The last one is the closed loop of the intent
owner. The intent owner comprehensively evaluates intent realization according to the feedback in
reports. If the intent is not realized or any conflict occurs, the intent owner takes measures
promptly to ensure that the intent can be realized continuously and effectively. For example, the
intent owner can solve intent conflicts or adjust intent expectations.

51 / 56
 Figure 33 Intent Fulfillment and Assurance
In an autonomous network, the intent processor is responsible for realizing the specific intent,
monitoring the state, and regularly generating intent reports for feedback to the upper layer,
starting from the underlying network element management layer. At each layer, intent reports
provided by the lower layer are used to assess the intent realization status, and adjustments and
optimizations are made if needed. The layer-by-layer bottom-up reporting and assessment
contribute to a closed-loop mechanism, which ensures that the intent is fully and efficiently
realized from the network element management layer to the business management layer. Through
continuous intent reporting and timely adjustments, the whole system can dynamically adapt to
changes to ensure continuous intent realization.

7.2.2 Large Model of Intent-Driven Network

With the boom of AI large models, China Mobile has built an AI large model system that
includes the basic large model and various industrial large models for government affairs,
industrial production, people's livelihood services, and communication networks. As the core of
general intelligence, the basic large model has the understanding generalization capability and
features output diversity and boundaryless knowledge. Industrial large models are based on the
basic big model and have integrated industry expertise, expert experience, and production data to
serve in industrial scenarios.
In an intent-driven autonomous network, it is necessary to further explore the integration of
general large models and industry intelligence. The full play to the exclusive advantages of large
models, including "cognition and understanding of panoramic network knowledge", "emergence
of capabilities for new scenarios", and "intent-driven interaction", will promote the evolution of
traditional networks to "intent-driven" networks and the digital and intelligent transformation and
upgrading of networks. At the same time, it is necessary to study the knowledge representation
and reasoning technologies of large models for IDNs to improve AI-based cognitive and
decision-making capabilities and meet the needs of intent-driven autonomous networks.
IDNs can translate users' service intents into a series of executable network tasks through
several steps, such as intent input, intent translation, conflict detection, intent execution, and intent
assurance. The networks can automatically perform conversion, verification, deployment,

52 / 56
configuration, and optimization according to the intents of users and operators to reach the target
network state. Exceptions can be solved automatically to ensure network reliability. However,
facing the differentiated needs of all service scenarios and complex network environments, current
IDNs still have some problems, such as inaccurate intent translation, inflexible intent policies, and
complicated intent management. Compared with conventional AI models, large models have
obvious advantages in intent understanding, reasoning, judgment, and decision-making. The
combination of large models and IDNs can effectively improve intent translation accuracy and
simplify the process of intent management and network O&M. The application of large models in
intent recognition and translation can improve the accuracy of user intent and demand recognition.
The application in policy matching and execution can intelligently match user demands and
network configuration solutions, realizing automatic network customization for differentiated user
demands. The application in network monitoring and maintenance can identify changes in the
monitored network environment, trigger network self-optimization, and update network
deployment or configuration.
Furthermore, the next-generation network architecture puts forward the requirement for
end-to-end multi-element joint scheduling and orchestration. Large models can help
comprehensively optimize end-to-end scheduling and improve network performance by enabling
end-to-end intent driving.

7.2.3 Intent-Driven Cross-Domain Autonomy

Automation of autonomous networks is based on the closed-loop mechanism and is a
feedback-driven process. The closed loop aims to achieve and maintain explicitly set intent goals
or automatically execute determined rules or policy instructions through perception, analysis,
decision-making, and execution. According to the timeliness requirements for closed-loop
processing at each layer of the autonomous network in different application scenarios, the closed
loop within the autonomous domain and the closed loop involving cross-layer interaction and
collaboration can be applied, respectively.
Intent-driven cross-domain network automation is a key research field. Further research and
exploration can be made on the application of the closed loop mechanism in the perception,
analysis, decision-making, and execution of autonomous networks, especially the way to achieve
and maintain explicitly set intent goals or automatically execute rules or policy instructions.
Moreover, it is required to pay attention to the timeliness requirements for closed-loop processing
at each layer of the autonomous network in different application scenarios, and explore the
collaboration between the closed loop within the autonomous domain and the closed loop
involving cross-layer interaction. In addition, it is necessary to study the details of how network
design and operations are dynamically adjusted to continuously optimize service delivery when
the expected intent goal is achieved in the autonomous domain.

53 / 56
 Abbreviations

English Full Name in English Chinese

Abbreviations
3GPP 3rd Generation Partnership Project 第三代合作伙伴计划
ADN Autonomous Driving Network 自动驾驶网络
AN Autonomous Network 自智网络
AI Artificial Intelligence 人工智能
CCSA China Communications Standards Association 中国通信标准化协会
CLL Cloud Leased Line 云专线
IBN Intent Based Network 基于意图的网络
IDN Intent Driven Network 意图驱动网络
IETF The Internet Engineering Task Force 国际互联网工程任务组
ICT Information and Communications Technology 信息与通信技术
ITU-T International Telecommunication 国际电信联盟电信标准
Union-Telecommunication Standardization Sector 化部门
NFV Network Functions Virtualization 网络功能虚拟化
RDF Resource Description Framework 资源描述框架
SDN Software Defined Network 软件定义网络
SLA Service Level Agreement 服务等级协议
TM Forum TeleManagement Forum 电信管理论坛
IDONP Intent-Driven Optical Network Platform 意图驱动光网络平台
UML Unified Modeling Language 统一建模语言

54 / 56
 References
[1] China Mobile Autonomous Network Whitepaper (2023)
[2] CCSA. Autonomous Network: System Architecture
[3] CCSA. Study on intent management technologies of autonomous networks
[4] TM Forum IG1253, Intent in Autonomous Networks.
[5] IRTF RFC 9316, Intent Classification.
[6] 3GPP TS 28.312, Intent driven management services for mobile networks.
[7] Huawei Autonomous Driving Network Whitepaper (2023)
[8] CIC and China Mobile CMRI: Whitepaper on Frontier Technologies of Autonomous Networks
[9] TM Forum Autonomous Network Whitepaper (3.0)
[10] Huawei Intent-Driven Network (IDN) Whitepaper
[11] China Mobile Whitepaper on Network Evolution Technology During 14th Five-Year Plan
[12] China Unicom Research Institute: Whitepaper on Intent-Based Network-Business
Collaborative Orchestration
[13] ZTE Autonomous Network Whitepaper

55 / 56
 Participants
Xidian University: Song Yanbo, Yang Chungang, Li Changle, Ouyang Ying, Mi Xinru, Li Tong
Beijing University of Posts and Telecommunications: Wang Jingyu, Guo Lingqi, Yang Hui, Yu
Tiankuo, Gong Shengye
University of Electronic Science and Technology of China: Li Longjiang, Shi Zhiping, Jiao Yifan,
Tang Zihui, Lei Weidong
Nanjing University of Aeronautics and Astronautics: Dong Chao, Zhu Xiaojun, Zhang Lei
China Mobile Research Institute: Deng Lingli, He Keguang, Ma Mengyuan
China Telecom: Guo Jianzhang, Zhang Jinsong, Li Xiangning, Li Zhen, Wang Dong
China Unicom: Gao Qian, Li Lu, Zhou Wei
ZTE: Ju Manchang, Wang Qilei, Xie Pengxiang

56 / 56