Skip to content
AsiaInfo's Network Ontology: Enabling AI to Understand, Judge, and Act on Networks
Operations 30 min read

AsiaInfo's Network Ontology: Enabling AI to Understand, Judge, and Act on Networks

AsiaInfo presents a network ontology architecture centered on objects, relations, rules, and actions, integrating RAG, knowledge graphs, and agents to give AI unified professional logic for telecom operations, validated by fault location and enterprise line survey cases showing minutes-level root-cause analysis and hour-to-minute survey acceleration.

AsiaInfo Technology: New Tech Exploration
AsiaInfo Technology: New Tech Exploration
AsiaInfo Technology: New Tech Exploration

Abstract

As large models and agent technologies see widespread adoption in network operations, single-scenario AI capabilities have matured. However, when AI scales into production systems, issues emerge: cross-agent professional cognitive inconsistency, duplicate knowledge construction, and poor rule reusability. Based on AsiaInfo's network ontology practice, this article proposes a network ontology architecture centered on four core elements—objects, relations, rules, and actions—to unify network professional logic and synergize existing AI capabilities. It details how the ontology enables AI to "understand the network, make judgments, and take actions," and validates the approach through two practical cases: intelligent link fault location and enterprise dedicated line resource survey, demonstrating value in reducing fault location time and improving solution generation efficiency. Finally, it discusses the evolution from scenario practice to capability precipitation, building reusable, operable digital assets.

1. Single-Scenario AI Works, But Production Reveals Four New Problems

Current network operations widely employ large models, Retrieval-Augmented Generation (RAG), knowledge graphs, and agents, achieving real results in fault analysis, complaint handling, network optimization, and knowledge Q&A. In a single scenario, prompt engineering, RAG, workflows, and expert rules enable AI to complete complex analysis and task execution. The core value at this stage is "retrieving knowledge and running processes."

As agents enter production, the focus shifts from "can AI answer?" to four production-grade problems:

Does AI acquire "relevant knowledge" or an accurate understanding of the current network state?

Do different agents share a consistent understanding of devices, links, services, and alarms?

Can expert rules and judgment logic be reused across scenarios instead of being rewritten into prompts and workflows repeatedly?

Can AI outputs feed into change, remediation, and closed-loop execution?

These converge on a single requirement: for high-level autonomous networks (L4/L5 per TM Forum AN7.0), network objects, relations, rules, and operational experience must be precipitated so AI gains stable, consistent, reusable professional cognition.

2. Ontology Does Not Replace RAG, Knowledge Graphs, or Agents—It Provides a Unified Professional Logic Framework

2.1 Distinct Roles of Each Technical Component

RAG : Solves "where to find knowledge"—retrieves from specifications, manuals, historical tickets, fault cases, expert experience; focuses on knowledge acquisition and context augmentation.

Knowledge Graph : Solves "what are the relationships between knowledge"—describes entities (devices, ports, links, services, customers, alarms) and their relations; focuses on knowledge association, query, and relationship expression.

Agent : Solves "how to complete the task"—task decomposition, tool invocation, result analysis, process execution; focuses on complex task automation.

Network Ontology : Solves "according to what professional logic to understand and use these capabilities"—organizes network professional knowledge into four core elements: objects, relations, rules, actions.

2.2 Four Elements Build a Unified Semantic Layer

Objects : What exists in the network—devices, ports, links, cells, customers, services, tickets, etc.; the basic entities of the ontology.

Relations : How objects connect—carry, connect, influence, depend, contain, associate; forming network topology and service chains.

Rules : How to judge under what conditions—fault rules, capacity rules, configuration rules, SLA rules; formalized expert experience.

Actions : What operations can be executed under what conditions—query, verify, simulate, dispatch, configure, change; connecting cognition to execution.

Thus AI moves from "finding knowledge" to "understanding current network → judging by rules → invoking capabilities to execute," achieving network understanding, judgment, and action.

3. Network Ontology Platform: A Five-Layer Capability Stack from Semantic Definition to Service Output

3.1 Unified Modeling and Runtime Service: Making Network Professional Knowledge Definable and Manageable

The platform provides unified construction, management, and runtime services, supporting unified definition and visual modeling of network entity objects, object relations, business logic/rules, and actions. It solves "how to build, who manages, how to run"—offering drag-and-drop modeling tools for network professionals to precipitate object definitions, relation constraints, and rule logic scattered across documents, code, and expert minds into structured, versioned network ontology assets.

3.2 Data Semantic Mapping: Connecting Static Ontology Definitions to Dynamic Network Data

Ontology definitions are only a semantic skeleton. The key to letting AI "read the network" is mapping ontology objects to actual production data. The platform builds data semantic mapping capabilities, connecting to data sharing platforms, resource management, and professional NMS systems, establishing mappings between ontology objects and resource, alarm, performance, configuration, topology, and ticket data. It supports entity association and dynamic instance state updates, enabling unified semantic queries across systems. Without mapping, the ontology is a castle in the air; with it, AI can query "current link bandwidth utilization" or "which services are on this port" under unified semantics instead of querying each system separately and manually aligning.

3.3 Query and Rule Reasoning Service: Assembling Traceable Network Context on Demand for Agent Tasks

The platform builds ontology query and rule reasoning services supporting object relation retrieval, business impact analysis, and rule constraint verification. When an agent receives a task, it doesn't traverse all network data; the ontology service extracts relevant objects, relations, reasoning rules, and business states per the task scenario, forming a structured, traceable network task context fed directly to the agent for analysis and decision. This "on-demand context extraction" ensures the agent gets real-time network state while avoiding context bloat from stuffing full data into the LLM.

3.4 Standardized Service Interfaces: Supporting Single-Agent and Multi-Agent Collaboration via API and MCP

Ontology capabilities are not bound to a specific agent but exposed via standard interfaces—Application Programming Interface (API) and Model Context Protocol (MCP). Single agents call ontology queries to understand network state; multi-agent collaboration aligns judgments on the same ontology semantics. Combined with the network operations skill library and existing production systems, the platform implements action constraint verification and controlled invocation—agents not only "see" the network but can "act" on it within rule-allowed bounds, with every step safety-checked by ontology rules.

3.5 Asset Governance: Continuous Operation Mechanism from Catalog, Versioning to Release and Rollback

Ontology is not static once built. The platform perfects asset catalog, version management, consistency verification, permission control, and release/rollback mechanisms. Based on application feedback, optimization suggestions are formed, validated, and used to update ontology models and rules, realizing continuous governance and shared reuse. This governance loop ensures the ontology model evolves with the network instead of becoming another outdated static knowledge base.

4. Resource Survey: From "Find Experts, Go On-Site" to AI Auto-Generating Solutions Based on Real Network Resources

4.1 Traditional Way: 7-Day Delivery Hard Constraint, 1-2 Weeks Manual Survey Becomes Business Bottleneck

Enterprise dedicated line business provides networking services for government and enterprises, covering complex urban sites and remote mountain access points. Taking a provincial telecom operator as example, 7-day delivery is a hard KPI, but resource verification and investment estimation routinely exceed one week. Frontline account managers cannot quickly verify line coverage or issue accurate quotes, easily losing project opportunities.

4.2 Ontology-Enhanced Solution Generation Process: Business Logic Decomposition Drives AI to Automate Full Process from Demand to Solution

After the customer provides address, bandwidth, and service level, existing resource agents can query resources, plan paths, and generate reports. The ontology-enhanced approach further unifies customer demand, resource capabilities, and business constraints. The system receives parameters (customer address/coordinates, service type, rate, number of circuits), uses relations among locations, equipment rooms, devices, ports, and links, and invokes the ontology's "business logic decomposition" capability to convert them into concrete resource query and verification conditions. Ontology logic orchestrates resource verification, planning, and estimation tools; AI then completes demand understanding, resource matching, rule verification, route planning, cost estimation, and solution recommendation end-to-end.

Taking a premium dedicated line as example (customer provides two addresses, Optical Transport Network (OTN) enterprise dedicated line access, 100M rate), the ontology-defined reasoning steps and business logic are:

Ontology service queries standard addresses and geographic coordinates for the two user-provided addresses.

Based on coordinates, queries equipment rooms containing OTN devices within 5 km.

For each equipment room, calls ontology service findGLDByZB to query fiber segment coverage rate, and findOtnUsing to query OTN utilization rate.

Ontology logic judges recommend/expand/new-build/abandon status based on OTN utilization and fiber coverage.

Lists all results, explaining candidate resource recommendation direction by coverage and utilization conditions. Example: Candidate A meets coverage and utilization (A: 85% / 65%); Candidate B exceeds utilization threshold (B: 75% / 82%), needs expansion assessment. Before actual access, still must verify resource instances, port capacity, and routing conditions.

Map displays resources and construction needs; report retains segmentation basis for solution review and subsequent construction assessment. Map shows planned route: customer demand mapped to actual line planning (demand address and construction segment address reflect different business stages), presenting start/end points, planned direction, and resource segments. Supports verifying plan locations (total route ~4.25 km, existing fiber ~2.10 km, new build ~2.15 km).

4.3 Closed-Loop Iteration: Survey–Correction–Re-Survey, Making Resource Data More Accurate with Use

Resource survey doesn't end with a one-time report. The system establishes a "survey–review–construction–correction" closed loop:

AI generates report with unique resource verification ID, attached with existing resource snapshot, routing planning basis, and cost estimation basis—full process traceable.

Survey personnel review report; if no objection, pass; if objection, provide evidence and correct resource data in-system; system regenerates report based on latest data, appending version number to original ID, preserving old report.

During construction, if deviation occurs, three handling cases: deviation due to inaccurate resource data → correct resource and trigger new survey; deviation due to unreasonable survey rules → adjust survey logic; deviation unproven but construction differs → correct construction per survey trace, control investment.

Every on-site survey and construction feedback becomes input for resource data correction and algorithm optimization, forming a virtuous "survey–construction–correction–feedback to survey" cycle.

4.4 Resource Survey: Engineering Effects of Ontology-Enhanced Intelligent Auto Solution Generation

Resource verification and budget quoting: compressed from 7+ days to hour-level.

Cost estimation vs. actual cost error: controlled below 20% after closed-loop iteration.

Covered business types: OTN premium dedicated line, IP Radio Access Network (IPRAN) dedicated line, Internet dedicated line, IP VPN, etc.

Frontline account managers shift from "waiting for professional feedback" to "self-service initial survey, rapid response to opportunities."

Why not just RAG? RAG can tell AI "what resources to check and what standards to follow for dedicated line provisioning," but generating a solution requires answering: Are there available devices within 5 km of this address? Which site has both free ports and free bandwidth? What fiber path is needed? How much budget overrun is reasonable? These answers are not in documents—they lie in current real network objects, topology relations, and resource states. The ontology automatically decomposes business demands into structured queries and rule reasoning on this real-time data.

5. Fault Location: From "Retrieving Experience" to Trusted Judgment Based on Real-Time Network State

5.1 From Manual Segment-by-Segment Troubleshooting to Minute-Level Root Cause Penetration

Traditional way: Link fault troubleshooting faces five pain points—alarm events, performance metrics, device info, port status, link topology stored separately, lacking unified correlation; fault diagnosis requires manual item-by-item check of port status, optical power, packet loss/error, protocol neighbors, traffic curves; NMS alarms may differ from real-time device status, false alarms cause invalid dispatches; single-sided faults, link flaps, logical faults, transmission relay faults hard to quickly differentiate; location conclusions, root cause categories, fault levels, and remediation suggestions inconsistent; troubleshooting time long.

Ontology-enhanced way: Build IP backbone "NE–port–link–alarm–performance" integrated resource twin business object and relation model; precipitate expert diagnosis experience, alarm correlation rules, and fault discrimination logic into system capabilities. Four concrete steps:

Build ontology model. Define five core object types: NE node, physical port, physical link, alarm event, performance metric. Establish inter-object associations—router contains physical ports, physical link connects A-end/Z-end ports, alarm event associates with device or port, performance metric belongs to link or port—forming an "alarm–performance–port–link–device" correlation network.

Bind reasoning logic. Bind standard judgment processes, single-sided fault judgment, link flap judgment, logical fault judgment, transmission relay fault, false alarm scenarios, and core reasoning logic to ontology objects to support automated analysis.

Layered diagnosis. Detect layer by layer—physical, data link, network/protocol—combined with rules for root cause judgment. Four core reasoning rules cover main fault types:

Link physical hardware fault: port operational state Down, corresponding link received optical power below threshold, and link hardware alarm exists.

Link quality degradation: port operational state Up, but packet loss >5% and Cyclic Redundancy Check (CRC) error packets continuously increasing.

Protocol layer fault: both ends port state Up, but Interior Gateway Protocol (IGP) neighbor state Down, and no physical layer alarm.

NMS data delay or false alarm: NMS has link Down alarm, but real-time device query shows port state Up.

Additionally covers six common exception branches: single-sided fault (one end Up one end Down, check optical module, fiber cross), link flap (frequent Up/Down switching, check optical power critical fluctuation), logical fault (physical Up but protocol Down, check configuration and neighbor parameters), transmission relay fault (both ports normal but middle interrupted, coordinate transmission to check fiber patch panel and WDM), false alarm (NMS vs. actual mismatch, exclude collection delay), performance degradation (link not down but quality declining, trend-based early warning).

Output standardized conclusions. Auto-generate fault location conclusion, root cause category, fault level, and remediation suggestion, assisting operations personnel in rapid handling.

5.2 Essential Difference Between Ontology-Enhanced Fault Diagnosis and Traditional Solutions

Why not just RAG? RAG can find "what causes a certain alarm type," but determining the actual cause of this fault requires simultaneous judgment of port status, received optical power, packet loss rate, CRC error trend, IGP neighbor state, and NMS alarm vs. real-time state consistency—this seven-dimensional combinatorial judgment is not in any single document; it exists in the cross-reasoning of current network real-time state and expert rules.

5.3 Engineering Effects of Ontology-Enhanced Fault Location

Root cause location timeliness: reduced from hours to minutes.

Root cause location accuracy: >95%.

Cross-domain fault propagation chain restoration completeness: >90%.

False alarm automatic identification, reducing invalid dispatches.

6. From Single-Point Intelligence to Scalable Autonomy: Precipitating Network Professional Cognition into Reusable Digital Assets

6.1 High-Value Scenarios First, Validate Technical Path with Business Value

Prioritize practice around real production problems: network resource survey, fault diagnosis, network change, complaint handling, network optimization, service assurance. First verify genuine improvements in efficiency, accuracy, and automation level, then expand to wider scope.

6.2 Extract Common Knowledge from Scenarios, Precipitate into Organizational Digital Assets

Continuously refine the four elements—objects, relations, rules, actions—gradually precipitating professional knowledge scattered in expert minds, prompts, code, and processes into reusable digital assets. This is not one-time modeling but continuous evolution with scenario iteration.

6.3 Multi-Agent Shared Professional Cognition, Forming Positive Evolution Loop

Future fault agents, complaint agents, optimization agents, resource agents, change agents still handle different tasks, but no longer need to each redefine what is a device, link, service bearing, applicable rule, or executable operation. Gradually form a continuous evolution mechanism: "scenario practice → professional knowledge precipitation → capability reuse → more scenarios."

The value of ontology is not to build another "concept platform," but in real network production to precipitate operators' network knowledge and expert experience into AI-era reusable, operable digital assets, supporting autonomous networks from single-point intelligence to scalable autonomy.

References

TM Forum. IG1255 Knowledge Management in Autonomous Networks v1.0.0 [S]. 2025.

ETSI. GS ENI 019 V4.1.1 — ENI Models [S]. 2026.

Zhou Q. Ontology-driven knowledge based autonomic management for telecommunication networks: theory, implementation, and applications [D]. Heriot-Watt University, 2021.

TM Forum. Autonomous Networks Implementation Guide (AN7.0) [R]. 2025.

AsiaInfo. Autonomous Network Ontology Practice [R]. 2026.

AsiaInfo. Lightweight Ontology Practice: Ontology-Based Big Data Platform Operations Agent [J]. New Technology Exploration, 2025.

Original Source

Signed-in readers can open the original source through BestHub's protected redirect.

Sign in to view source
Republication Notice

This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactadmin@besthub.devand we will review it promptly.

AI agentsRAGKnowledge GraphAIOpsAutonomous NetworksTelecom OperationsAsiaInfoNetwork Ontology
Reader feedback

How this landed with the community

Sign in to like

Rate this article

Was this worth your time?

Sign in to rate
Discussion

0 Comments

Thoughtful readers leave field notes, pushback, and hard-won operational detail here.