| 摘要: |
| 数字孪生网络（Digital Twins Network，DTN）为移动通信网络部署与优化提供了物理空间和数字空间联动的新范式，而高精度、可泛化的室内无线传播建模是DTN实现无线环境精准映射的关键前提。然而现有室内无线传播模型存在精度不足、泛化性弱、难以适配动态环境的问题，无法满足DTN无线环境精准映射需求。为此，提出了一种基于全卷积网络（Fully Convolutional Networks，FCN）的编解码器架构模型ResTrans-Net。该模型的编码器融合了残差网络（ResNet）和Transformer，以增强局部特征提取与全局信息分析能力，进而提升预测精度。为应对环境动态变化带来的挑战，引入了基于干扰挖掘的自适应校准方法，使模型能够适应新环境并保持稳健性能。实验结果表明，ResTrans-Net 在未知室内场景下路径损耗预测的均方根误差（Root Mean Square Error，RMSE）低至374 dB，较主流SDU-Net、U-Net模型误差分别降低 20.9%、34.8%；经36个点位自适应校准后，测试集RMSE可进一步降至2.69 dB。 |
| 关键词: 数字孪生网络(DTN) 室内传播模型 全卷积网络 模型校准 迁移学习 |
| DOI：10.20079/j.issn.1001-893x.260109002 |
|
| 基金项目:国家科技重大专项（2026ZD1307800）；“北京邮电大学-中国联通联合创新中心”联合研发项目（2025-STHZ-BJYDDX-007） |
|
| Indoor Propagation Model with Adaptive Calibration for Digital Twins Network |
| WANG Yaning,WANG Jinshi,CHENG Yihang,PENG Tao,WANG Wenbo |
| (Key Laboratory of Universal Wireless Communication，Ministry of Education，Beijing University of Posts and Telecommunications，Beijing 100876，China) |
| Abstract: |
| Digital twins network(DTN) provides a new paradigm for the linkage between physical space and digital space in the deployment and optimization of mobile communication networks，and high-precision，generalizable indoor wireless propagation modeling is a key prerequisite for DTN to achieve accurate mapping of the wireless environment.However，existing indoor wireless propagation models have the defects of insufficient accuracy，weak generalization ability and poor adaptability to dynamic environments，which cannot meet the requirement of DTN for accurate mapping of wireless environment.Accordingly，an encoder-decoder architecture model based on fully convolutional networks(FCN)，named ResTrans-Net，is proposed.The encoder of this model integrates residual network(ResNet) and Transformer to enhance the capabilities of local feature extraction and global information analysis，thereby improving prediction accuracy.To address the challenges brought by dynamic environmental changes，an adaptive calibration method based on interference mining is introduced，enabling the model to adapt to new environments and maintain robust performance.Experimental results show that the proposed ResTrans-Net achieves a root mean square error（RMSE） as low as 3.74 dB for path loss prediction in unseen indoor scenarios，corresponding to a 20.9% and 34.8% relative reduction in RMSE compared with the mainstream SDU-Net and U-Net models，respectively.After adaptive calibration with 36 calibration points，the RMSE on the test set of the same unseen scenarios is further reduced to 2.69 dB. |
| Key words: digital twins network(DTN) indoor propagation model fully convolutional networks model calibration transfer learning |
《电讯技术》编辑部 