信道通用预训练大模型赋能数字孪生信道:原理与实践  

Channel General Pre-trained Large Model Empowering Digital Twin Channel:Theory and Practice

作　　者：张建华 史廉正 于力 黎明月 田磊 王启星 刘光毅 ZHANG Jianhua;SHI Lianzheng;YU Li;LI Mingyue;TIAN Lei;WANG Qixing;LIU Guangyi(School of Information and Communication Engineering,Beijing University of Posts and Telecommunications,Beijing 100876,China;School of Electronic Engineering,Beijing University of Posts and Telecommunications,Beijing 100876,China;China Mobile Future Institute,Beijing 100032,China)

机构地区：[1]北京邮电大学信息与通信工程学院,北京100876 [2]北京邮电大学电子工程学院,北京100876 [3]中国移动未来研究院,北京100032

出　　处：《无线电工程》2025年第4期679-686,共8页Radio Engineering

基　　金：国家重点研发计划(2023YFB2904805);国家自然科学基金(62401084)。

摘　　要：未来第六代移动通信系统(6G)以提供超可靠、智能化的网络连接,实现动态环境下的万物互联。数字孪生信道(Digital Twin Channel,DTC)作为支持6G网络智能化的关键技术,在数字世界中在线构建高保真孪生信道,助力网络主动适应与精准决策。传统的AI小模型通常仅针对特定任务或特定场景,面对6G复杂、高动态的无线信道存在巨大挑战,难以全面满足DTC的需求。大语言模型(Large Language Model,LLM)因在多模态特征融合、高维数据建模等方面具有强大能力,有望克服上述挑战。首次将DTC与LLM框架相结合,构建信道通用预训练大模型ChannelGPT,以应对6G无线信道的巨大挑战。具体来说,ChannelGPT设计为3个核心层,数据处理层构建大规模环境与信道数据集,算法模型层充分挖掘无线环境信息与信道特性之间的匹配映射,功能应用层有望支持多任务的联合优化。仿真结果表明,ChannelGPT在预测精度、多模态融合能力方面相较于小模型方法具有明显优势,为实现DTC提供了有潜力的手段。The future 6G aims to provide ultra-reliable,intelligent network connectivity and realize the internet of everything in dynamic environments.Digital Twin Channel(DTC),as a key technology supporting the intelligence of 6G network,constructs high-fidelity twin channels online in the digital world to assist with proactive network adaptation and precise decision-making.However,traditional AI small models are usually tailored to specific tasks or scenarios.These models face significant challenges in predicting complex,and highly dynamic wireless channels of 6G,which struggle to fully meet the requirements of DTC.Large Language Model(LLM),with its powerful capabilities in multimodal feature fusion,high-dimensional data modeling and so on,have the potential to overcome these challenges.The DTC framework is integrated with LLM for the first time in this research,proposing a channel general pre-trained large model(ChannelGPT) to tackle the substantial challenges of 6G wireless channels.Specifically,ChannelGPT is designed with three core layers.The data processing layer constructs large-scale environment and channel datasets.The algorithm model layer fully explores the mapping between wireless environment information and channel characteristics.The functional application layer is expected to support multi-task joint optimization.Simulation results demonstrate that ChannelGPT outperforms small models in prediction accuracy and multimodal fusion capabilities,providing a promising approach for realizing DTC.

关 键 词：大语言模型 信道预测 数字孪生信道 深度学习 

分 类 号：TN929.5[电子电信—通信与信息系统]