PaperlayerView exact revision on arXiv

Source-linked AI summary

LLM-Enabled NWDAF: A Step Toward AI-Native 6G Network Intelligence

Henok Daniel, Omar Alhussein, Cheng Li, Jie Liang, Ernesto Damiani

arXiv:2606.11877v1cs.NI

TL;DR

Open-source 5G networks lack native intent-based interfaces, limiting accessible network analytics and control. This paper extends Free5GC with an LLM-powered NWDAF, achieving 98.43% intent-classification accuracy across seven operational categories and improving operator usability.

  • Problem

    Open-source 5G networks lack native interfaces that translate high-level operator intents into analytics-driven actions.

  • Method

    The paper extends Free5GC with an event-driven NWDAF and a RAG-based LLM interface that maps natural-language intents to analytics queries or subscription commands.

  • Results

    98.43% accuracy was achieved for intent classification across seven operational categories.

  • Takeaways & Limitations

    The implementation validates natural-language interaction with event-driven analytics and improves NWDAF usability for operators.

  • Takeaways & Limitations

    The current LLM interface is limited to intent classification and basic analytics queries, leaving advanced automation and policy enforcement for future development.

Abstract

from arXiv · show

The Network Data Analytics Function (NWDAF) is central to enabling zero-touch network management in fifth-generation (5G) networks by supporting real-time analytics and closed-loop automation. Despite its critical role, open-source NWDAF implementations remain limited in scope and accessibility. In this paper, we develop an open-source NWDAF, compatible with the open-source core network Free5GC, that collects network data via subscriptions to Network Functions (NFs), and also includes an integrated Large Language Model (LLM) interface that enables natural language interaction with human operators. The interface processes user intents, encodes them using a semantic embedding model, and maps them to one of seven predefined intent categories to trigger analytics queries or event subscription commands. This architecture abstracts the complexity of traditional interfaces, allowing non-expert users to manage network analytics and subscriptions with ease. The system supports Access and Management Function (AMF) and Session Management Function (SMF) event subscriptions, real-time monitoring, and analytics retrieval via Prometheus, all accessible through a conversational interface. By bridging AI-driven intent recognition with standardized network analytics, our implementation enhances operator usability and provides a foundation towards AI-native 6G networks. The source code and datasets generated during the current study are available in the github repository, https://github.com/HenokDanielbfg/testbed.

Introduction

The introduction presents NWDAF as a foundation for data-driven, closed-loop 5G automation and motivates an LLM-powered intent interface that makes network analytics and control more accessible. The proposed open-source Free5GC extension supports event-driven monitoring, intent-based queries and subscriptions, and achieves 98.43% intent-classification accuracy across seven operational categories.

  • Motivation: 3GPP Release 15 introduced NWDAF to aggregate and analyze data from multiple Network Functions for closed-loop 5G automation.NWDAF enables analytics exchange across NFs and supports zero-touch network and service management.
  • Motivation: Open-source 5G networks lack native interfaces for intent-based networking, which translates operator objectives into analytics and control actions.IBN shifts network management from procedural configuration toward declarative objectives expressed at a higher level.
  • Contributions: The proposed system extends Free5GC with an LLM-powered NWDAF that collects real-time UE registration, session setup, and handover events through NF subscriptions.The implementation supports AMF and SMF event subscriptions and real-time notification handling.
  • Contributions: User prompts are semantically matched against predefined intents through a Retrieval-Augmented Generation module, then executed as analytics queries or subscription commands.The classified intent is translated into internal representations such as REST API calls.
  • Results: 98.43% accuracy was achieved for intent classification across seven operational categories.The system also generates network event datasets from AMF and SMF using a custom UERANSIM mobility model to trigger mobility-related events.
  • Contributions: The codebase is released as open source to support future research on NWDAF capabilities in open 5G testbeds.The interface is intended to lower barriers for researchers, students, and operators exploring network behavior without custom code or detailed NF-interface knowledge.

Related Work

Related work examines how 5G service-based architecture and virtualized RAN have shaped research on core networks, UE behavior, and data analytics. It focuses on 5G core and RAN implementation, NWDAF integration, intent-based networking, and mobility modeling.

  • 5G service-based architecture and RAN virtualization have driven research into 5G core development, UE behavior, and data analytics.
  • The reviewed work covers 5G core and RAN implementation alongside NWDAF integration.
  • It also addresses intent-based networking in 5G and mobility modeling.

5G Core and RAN Implementation

Existing 5G platforms and NWDAF studies provide open-source implementations, simulated analytics, and architectural foundations, while recent work increasingly combines intent-based networking with AI and LLMs. Building on these directions, the paper unifies LLM intent recognition with real-time NWDAF analytics and subscription orchestration.

  • Open-Source 5G Platforms: Open5GS offers a flexible, scalable 5GC implementation aligned with 3GPP Release 17, while UERANSIM supports 5G technology development and testing.These projects contribute to open-source 5G implementation and experimentation.
  • Prior NWDAF Implementations: Prior NWDAF research spans conceptual federated architectures, simulated protocol classification, and Open5GS-UERANSIM datasets for analyzing control-plane signaling and NF placement.The federated design uses leaf and root NWDAFs, while another study reports approximately 65% average accuracy from Random Forest and Decision Tree models.
  • Prior NWDAF Implementations: A Free5GC-compatible NWDAF aligns with Free5GC NF architecture, separates model training and analytics delivery functions, registers with the NRF, and evaluates analytics using MNIST.Its integration with other NFs for live data collection and analytics remains unfinished in the supplied passage.
  • Intent-Based Networking and LLMs: Research on intent-based networking translates high-level objectives or natural-language requests into enforceable policies for automated 5G configuration, optimization, slicing, positioning, and orchestration.These systems emphasize closed-loop control, low-latency translation, and usability for non-expert stakeholders.
  • Intent-Based Networking and LLMs: Recent work combines IBN with NWDAF analytics and LLM-based intent parsing, motivating an end-to-end framework that integrates intent recognition, real-time analytics, and subscription orchestration.The combined approach supports proactive service assurance, including traffic forecasting, anomaly detection, slice scaling, and resource allocation.

5G System Architecture

The 5G architecture combines a service-based core, an NR radio access network, and standards-compliant user equipment. Its service-oriented interfaces and separated signaling and user-plane paths support dynamic network-function interaction and scalable communication.

  • Service-Based Core: 5G’s service-based architecture exposes network-function capabilities as discrete services through standardized, typically RESTful, Service-Based Interfaces.Stateless, endpoint-agnostic communication enables network functions to dynamically discover and invoke one another.
  • Service-Based Core: The Network Repository Function registers active network functions and service profiles, enabling dynamic selection, load balancing, horizontal scaling, and failover.For example, the AMF queries the NRF to select an available SMF instance and establish a transient connection.
  • NR RAN: In the NR RAN, the gNodeB integrates control-plane functions for RRC, mobility, and QoS negotiation with user-plane processing across PDCP, RLC, MAC, and PHY layers.This integration follows 3GPP specifications.
  • NG Connectivity: The NG interface separates N2 signaling, using NGAP for registration and mobility control, from N3 user-plane traffic tunneled by GTP-U toward the UPF.This split permits independent scaling and optimization of signaling and data transport paths.
  • 5G Baseline: SBA-driven core, NR-based RAN, and standards-compliant UE together form the baseline architecture for 5G deployments.The UE stack includes NAS for core registration, authentication, and session management, plus AS for radio protocols and RAN interaction.

Implementation

The implementation combines an open-source 5G testbed with an NWDAF that collects NF data, supports AMF/SMF event subscriptions, and adds a conversational LLM interface for analytics and subscription management.

  • 5G testbed: The testbed integrates a 5G Core Network, Radio Access Network, and User Equipment simulation using open-source software.The core network is implemented with Free5GC, which provides the 5G core network functions.
  • Event subscriptions: AMF and SMF expose RESTful APIs for subscriptions to registration, mobility, connectivity, and session-management events.AMF subscriptions cover UE registration, deregistration, handover updates, and connectivity-state changes; SMF subscriptions cover PDU session establishment and modifications.
  • NWDAF architecture: The NWDAF collects data from other NFs and processes it into network insights, extending a Free5GC-based design with service, data collection, prediction, and LLM modules.The architecture includes an NWDAF component that supports NF data collection and insight generation for interested parties.
  • LLM interface: The conversational interface translates natural-language operator requests into Prometheus analytics queries or NWDAF REST subscription commands using RAG-based intent matching.It combines intent embeddings, semantic similarity matching, function routing, and natural-language response generation to support flexible conversational interaction.

Data Generation and Analysis

The two-week testbed evaluation collected mobility and network-state data, showing predictable handover behavior and temporal variation in UE activity. The LLM interface supported analytics-oriented queries, with embedding-based intent classification outperforming direct LLM classification and correctly extracting network information while sometimes adding unnecessary text or speculation.

  • Testbed evaluation: The two-week experiment used four virtual gNodeBs in a square formation, each with a distinct TAC and 120-unit simulated coverage radius.The coverage radius corresponded to a signal strength threshold of -120 dBm.
  • Testbed evaluation: All UE registration and deregistration attempts succeeded, while average active-state durations ranged from 100 to 102 minutes.The collected data covered UE registration, state transitions, and handovers.
  • UE activity and mobility: Connected UE counts peaked around 11:00 and between 14:00 and 17:00, while the longest continuous UE activity session lasted 9000 seconds (2.5 hours).Some hours had no active UEs, whereas others showed moderate load.
  • Handover prediction: 80.65% accuracy was achieved by Gradient Boosting, followed by Random Forest at 80.24%, Decision Tree at 80.11%, and K-Nearest Neighbors at 79.03%.The dataset was split into 70% training and 30% testing sets.
  • LLM interface evaluation: 98.43% of prompts were correctly classified by text-embedding-ada-002, compared with 96.86% for all-MiniLM-L6-v2, 89.5% for GPT-4o, and 55.1% for GPT-4o-mini.The embedding approach used cosine similarity against a predefined database of validated intents.

Conclusion

The paper presents a software-based 5G testbed combining Free5GC, an LLM-enhanced NWDAF, and UERANSIM, with event subscriptions and automated RAG-based network analysis. It also identifies future extensions toward autonomous data retrieval, policy enforcement, and intelligent self-operating networks.

  • Conclusion: The testbed integrates Free5GC with a fully functional LLM-enhanced NWDAF and UERANSIM with a custom mobility model.The implementation includes NWDAF event-subscription mechanisms and automated network analysis through a RAG-based LLM-powered NWDAF.
  • Conclusion: Experiments validated the system’s ability to capture network events, support event-driven analytics, and predict UE behavior.
  • Conclusion: Future LLM-interface development could enable autonomous supplementary-data requests from relevant NFs to provide more dynamic, context-aware responses.This would extend the interface beyond intent classification and basic analytics queries.
  • Conclusion: More advanced automation and policy enforcement would move the system toward fully intelligent and self-operating network control.The proposed direction would bridge human-friendly commands and low-level network control tasks through NWDAF service functions based on user intents.

Additional information

The authors report no competing interests and no funding received.

  • The authors declare no competing interests.
  • No funding was received.

Appendix

The appendix describes how NWDAF collects AMF and SMF event reports, exposes them as Prometheus metrics, and provides representative records for three network-analytics datasets. These metrics support monitoring, session and mobility analysis, and registration-state evaluation.

  • Data collection and exposure: AMF and SMF event notifications are delivered through HTTP callbacks, parsed by NWDAF, and exposed as Prometheus metrics for real-time monitoring and querying.Subscribed events include UE registration, PDU session changes, and handovers.
  • Prometheus metrics: Active_UEs tracks connected-device counts and durations at each timestamp for session-management analysis and capacity planning.The appendix identifies this as Table 2.
  • Prometheus metrics: UE_location_report records each UE’s SUPI, serving NrCellId, and TAC to support mobility tracking and handover prediction.The appendix identifies this as Table 3.
  • Prometheus metrics: UE_registration_state logs per-UE active/inactive transitions with timestamps and state durations for evaluating engagement and optimizing registration transitions.The appendix identifies this as Table 4.
Loading 2606.11877v1…