Applying Large Language Models to Wireless Network Intelligent Operations: Opportunities, Challenges, and Platform Construction

The rapid development of 5G and upcoming 6G networks has made wireless architectures increasingly complex, creating a need for advanced AI solutions to address issues such as coverage, resource management, interference, and multi‑scenario optimization. Large language models (LLMs) like ChatGPT, GPT‑4, and domestic counterparts (e.g., Huawei PanGu, Baidu Wenxin) offer strong knowledge extraction, multimodal capabilities, and self‑learning, making them promising for telecom intelligent operation.

Key challenges for applying AI to wireless networks include:

Absence of standardized, publicly available datasets, hindering reproducibility.

Difficulty integrating NLP/vision AI techniques with domain‑specific wireless data and expert knowledge.

Varying communication scenarios (indoor, outdoor, high‑speed rail) and limited compute resources, requiring models that can operate efficiently across diverse conditions.

Lack of systematic analysis on performance bounds, reliability, and cost considerations for AI‑driven solutions.

To address these issues, the article proposes a unified intelligent‑operation platform built on LLMs. The platform consists of five main components:

Log data preprocessing (cleaning, feature extraction, normalization).

Vector database for efficient storage and retrieval of processed logs.

Prompt templates designed with chain‑of‑thought (CoT) techniques to guide model reasoning.

Domain‑specific knowledge graph constructed from expert experience.

LLM backend (e.g., ChatGLM2‑6B) fine‑tuned via P‑Tuning v2 for tasks such as root‑cause analysis and anomaly detection.

The platform architecture is illustrated in the following diagram:

Experimental validation used the ChatGLM2‑6B model with prompts and responses such as:

{
  "prompt": "告警数据有2条，第0条数据中，子原因是人为操作，告警项是RHUB不在位，故障类型是规划RHUB，小区号是NoCELL，发生的时间顺序是3374,第1条数据中，子原因是链路异常，告警项是射频单元不在位告警，故障类型是规划RRU，小区号是NoCELL，发生的时间顺序是20,",
  "response": "这条告警数据的根因是链路异常。"
}

Results on single‑task fine‑tuning showed high accuracy (Root‑Cause Analysis: 97.7% and 90%; Anomaly Detection: 87.4%). A mixed‑task model trained on both tasks achieved comparable performance (Root‑Cause Analysis: 84.4%; Anomaly Detection: 87.1%). These findings demonstrate that LLMs can outperform traditional methods and support a unified multi‑task framework.

Future work includes scaling to larger models, further fine‑tuning, expanding multi‑task evaluations, and integrating a Long‑Chain AI agent with a telecom‑specific knowledge base to achieve fully automated intelligent operation.

References: 6G Physical‑Layer AI Whitepaper, China Mobile Research Institute, 2022. 6G Wireless In‑born AI Architecture Whitepaper, China Mobile Research Institute, 2022. AIGC (GPT‑4) Empowering the Telecom Industry, AsiaInfo & Tsinghua University, 2023.