Structured Signaling Large Model: Facilitating the Intelligent Revolution at the Protocol Level

5G-Advanced (5G-A) is accelerating the intelligent transformation of high-value industries such as cloud gaming, the industrial metaverse, telemedicine, and vehicle-to-everything (V2X), raising network quality requirements from "usable" to "extremely reliable." However, the complexity of signaling interactions has grown exponentially, with a single user generating tens of thousands interactions daily. Traditional cross-domain data record (XDR) analysis suffers from cell loss (>30%) and limited scenario coverage (60%), forcing O&M personnel to manually parse raw signaling—often taking over eight hours on average. Additionally, complex faults (e.g., timing conflicts and protocol compatibility issues) are difficult to localize, impacting complaint resolution and network optimization.

ZTE has introduced a structured signaling large model that overcomes the barriers to intelligent protocol understanding, shifting from "manual signaling decoding" to "model-based protocol cognition." Leveraging machine cognition and end-to-end reasoning, the solution redefines the signaling analysis paradigm, significantly enhancing operators' intelligent O&M capabilities. Compared with traditional methods, it improves signaling analysis efficiency by 80%, raises fault localization accuracy to over 95%, and enables operators to deeply participate in scenario-specific optimization during the fine-tuning phase—helping build an intelligent user experience assurance system for the 5G-A era.

Innovative Signaling Analysis Solution Based on Structured Large Model

The solution establishes an end-to-end intelligent signaling analysis system through a collaborative architecture comprising data, model, and service layers, enabling a leap from raw signaling parsing to intelligent reasoning analysis.

Three-Layer Architecture Design

The architecture is shown in Fig. 1:

• Data layer: Supports efficient encoding of raw signaling, covering over 100 types of network element interfaces and related protocols, including 5GC, EPC, and IMS, ensuring the integrity and consistency of signaling data.
• Model layer: Integrates structured signaling large models and large language models (LLMs), providing pre-trained foundational models, chain-of-thought fine-tuned models, and intelligent reasoning capabilities. This enables accurate understanding of complex signaling logic and enhances fault diagnosis and anomaly analysis.
• Service layer: Incorporates real-world business scenarios to provide signaling visualization, anomaly detection, automated analysis reports, and intelligent Q&A, empowering operators to achieve intelligent network O&M.

Structured Signaling Large Model: Evolving From Rule-Driven to Intelligent Cognition

The solution uses structured large models to learn original signaling interaction patterns and integrates a chain-of-thought mechanism based on expert experience, enabling the model to autonomously parse signaling.

Signaling is essentially the "language" between network devices, characterized by fixed formats and low information dispersion, making it more suitable for LLM-based methods than natural language.

As a result, by deeply learning network protocol features, the model achieves end-to-end signaling parsing, protocol conflict detection, and abnormal pattern recognition.

Three Core Models Enabling End-to-End Intelligent Signaling Analysis

The solution establishes a comprehensive intelligent signaling analysis architecture through the collaborative optimization of encoding models, projection models, and domain decoding models. It also provides operators with opportunities for deep participation during the fine-tuning phase to ensure precise adaptation of the models to real-world network environments.

• Signaling encoding model: Built on an improved Transformer and a multi-level attention mechanism, it analyzes the byte-level content and hierarchical structures of signaling protocols, providing accurate semantic representations of signaling.
• Projection model: Constructs a mapping between the signaling semantic space and the business rule space, unifies data in different protocol formats, achieves automatic feature alignment, and reduces adaptation complexity.
• Domain decoding model: Enhances signaling reasoning capabilities and supports various intelligent O&M applications by embedding 3GPP protocol standards and integrating signaling processes, messages, and business scenarios.

Applying RAG for Precise Signaling Knowledge Reasoning

By integrating retrieval-augmented generation (RAG), the solution significantly improves the accuracy and efficiency of signaling parsing:

• Retrieval optimization: Introduces pre-retrieval routing, query rewriting, index optimization, and re-ranking techniques to greatly enhance the efficiency and relevance of signaling data retrieval.
• Reducing hallucinations in large models: Combines foundation models with external knowledge sources to optimize the information generation process, ensuring the accuracy and interpretability of reasoning results.

Application Empowerment: Complaint Analysis Agent

As 5G evolves toward 5G-A and future 6G, network services are becoming increasingly diverse, making user experience assurance a core demand. Faced with the complexity of new services and scenarios, traditional user complaint handling models struggle to provide rapid response and precise issue identification.

ZTE has developed a complaint analysis agent powered by the structured signaling large model, introducing a digital employee—the "user complaint handling expert"—to achieve intelligent and automated end-to-end signaling analysis, helping operators improve efficiency and reduce costs.

• Intelligent complaint Q&A: Lower the threshold for signaling analysis and improving response efficiency

Complaint-related signaling data is vast and complex, with intricate protocol flows. Manual analysis relies heavily on expert experience, and fragmented knowledge leads to low processing efficiency. By leveraging a structured signaling large model, ZTE has developed an intelligent Q&A engine that provides signaling analysts with immediate and accurate knowledge support through human-machine collaborative interaction and historical knowledge accumulation. This lowers the barrier to  signaling processing and enhances analysis efficiency.

• Automated complaint signaling analysis: Accurately pinpoint anomalies and swiftly provide solutions

Users can click on the agent entry point or directly input their issue intent to trigger the automated complaint analysis process. The agent enables automatic identification of abnormal signaling and intelligent screening, parses and visualizes signaling processes for intuitive root cause analysis, provides standardized signaling explanations to lower technical barriers, and offers handling suggestions and failure cases to support precise decision-making.

• Intelligent complaint analysis report generation: Free up O&M resources and improve service quality

The solution integrates RAG technology and consolidates multi-domain and multi-type data to generate comprehensive and accurate complaint analysis reports. Report generation time is significantly reduced, minimizing manual effort and enhancing the standardization and automation of the complaint handling process.

In 2025, ZTE completed key technology verification of an intelligent complaint analysis agent based on a structured signaling large model in Jiangsu, marking a milestone in intelligent complaint handling. Moving forward, ZTE will strengthen its focus on 5G-A networks, and launch intelligent complaint resolution solutions for both ToC (individual users) and ToB (enterprise users), helping operators build a more efficient, accurate, and intelligent complaint management system.

As a leading global provider of telecommunications equipment and network solutions, ZTE will continue to drive innovation and advance digital and intelligent  networks, delivering superior network experiences for customers worldwide.