LLM-Based Multi-Agent System for Intelligent O&M of Independent Private 5G Networks

With the rapid adoption of independent private 5G networks in industrial digital transformation, tensions between complex network O&M and increasingly refined industry demands are becoming more apparent. The traditional O&M model, which relies on manual experience, often suffers from delayed responses and high costs, making it difficult to meet the stringent requirements for network stability, low latency, and high reliability on industrial sites.

Based on the UniEngine integrated computing and network platform, ZTE proposes an intelligent O&M solution for independent private 5G networks, powered by a large language model (LLM)-based multi-agent system. By integrating endogenous intelligence with industry-specific customization capabilities, the solution offers a new approach to the construction and operation of private 5G networks.

Endogenous Intelligence: Building a Digital Brain with a ToB Agent Matrix

The core technology behind intelligent O&M for independent private networks is an endogenous intelligent architecture that enables an intelligent O&M system for industrial sites through a three-layered framework (Fig. 1).

At the bottom layer, the GPU-powered edge computing engine, based on lightweight edge servers, enables local AI inference and real-time data processing, significantly reducing the dependence on cloud computing power, ensuring that high-sensitivity data remains on site, and enhancing data security.

At the middle layer, serving as the agent’s "cognition center", ZTE's Nebula large model provides natural language understanding, multimodal data integration, and decision-making and inference capabilities, enabling the agent to meet the diversified requirements of complex scenarios.

At the top layer, a ToB agent matrix, built upon the Nebula large model, forms a closed-loop O&M system:

• The Q&A agent responds to customer and work order queries in real time through multilingual interactions and semantic correction, replacing traditional manual customer service.
• The monitoring agent automatically establishes hierarchical network connections based on customer requirements, analyzes network KPI (e.g., delay, bandwidth usage), device status, and environmental sensor data in real time, and implements abnormal behavior prediction and visual early warning.
• When detecting network connectivity or delay issues, the assurance agent automatically checks parameters, collects service logs, and diagnoses and analyzes services and networks based on decision logic.

Through "user-network-cloud" collaborative deployment, this architecture deeply integrates network-layer and service-layer O&M, enabling an end-to-end intelligent O&M closed-loop for industrial sites. With real-time responsiveness and autonomous decision-making capabilities, it provides "O&M-free" support for 5G industrial site networks, enabling vertical industries to make an intelligent leap in both network and business operations.

Capability Expansion: Building Expert Agents by Injecting Industry Knowledge

Facing the complex and differentiated O&M challenges in vertical industries such as energy, manufacturing, and other industrial sectors, ZTE’s intelligent O&M large model for independent private networks significantly enhances the domain adaptability and decision-making precision of agents through an industry knowledge injection mechanism. This mechanism leverages the intelligent digital operation service (IDOS) of the UniEngine to build three core capabilities:

• Customized knowledge base: Users can inject structured and unstructured data, such as device manuals, O&M SOP, and historical work orders, into the Nebula large model to create a dedicated knowledge base. For example, a steel company can upload parameters of the blast furnace cooling system, allowing the assurance agent to accurately identify the correlation between water temperature anomaly and equipment aging, improving root cause identification efficiency.
• Self-orchestrated knowledge graph: Through the IDOS platform, users can construct device topologies, software-hardware mapping relationships, and data service flow diagrams in the production domain. When a fault occurs, the ToB agent dynamically reasons over the self-orchestrated knowledge graph, simulates the fault propagation path, and quickly identifies the root cause.
• Real-time enterprise database access: With authorized access to enterprise databases (such as ERP and MES), the agent can retrieve the latest service rules and device records in scenarios such as work order processing and fault diagnosis, ensuring that decision-making and service updates are synchronized. For example, a power company can embed the latest inspection standards into the Q&A agent to enable standardized responses.

These capabilities enable the ToB agent to evolve from a "general assistant" to a "domain expert", significantly improving the accuracy of O&M decision-making and expanding scenario coverage.

Capability Opening: Building an Agent Platform to Let Customers Act as AI Product Managers

To unlock the full innovation potential of the industry, ZTE has launched an agent construction platform centered on low-code development, breaking through the limitations of traditional O&M tools and enabling customers to flexibly construct professional agents tailored to their specific needs. The platform provides a graphical workflow orchestration function, enabling users to combine the input, processing logic, and output of an agent in a modular manner with simple drag-and-drop operations, quickly building a customized O&M process. It supports API plug-in integration, allowing users to encapsulate the API capabilities of existing systems as standard plug-ins and seamless embed them through parameter configuration for seamless interconnection with the service system. The platform also features a built-in industrial agent template library, covering preset solutions for industrial, energy, and manufacturing sectors. Users can invoke these templates with one click or further develop them, significantly lowering the threshold for innovation. This platform-based capability enables enterprises to shift from using tools passively to actively defining intelligence, accelerating the deep integration of 5G private networks and industrial scenarios.

Future Outlook: From "O&M-Free" to "Self-Evolution"

The ultimate goal of intelligent O&M for independent private networks is to build a self-evolving industrial digital intelligent agent. By continuously learning from industry data, user feedback, and network operation status, the agent will autonomously iterate model parameters and optimize O&M policies, ultimately establishing a new paradigm of network as a service (NaaS). ZTE is working with global industry partners to promote the upgrade of 5G private networks from mere connectivity infrastructure to intelligent production hubs, injecting continuous innovation momentum into Industry 4.0.

The LLM-based multi-agent system for intelligent O&M of independent private 5G networks represents not only a technological breakthrough, but also a paradigm shift in O&M. Built on endogenous intelligence and empowered by knowledge injection and agent platforms, it enables industrial 5G networks to achieve truly unattended operation and autonomous decision-making.