| 摘要: |
| 提出了一种结合时间注意力机制与改进的2D全卷积神经网络-门控循环单元（2D-Fully Convolutional Neural Networks - Gated Recurrent Unit,2DFCNN-GRU）模型的信道状态信息（Channel State Information,CSI）预测方法，通过特征金字塔网络（Feature Pyramid Network,FPN）提取时变信道特征，并结合U-Net架构进行特征融合。利用GRU进行离线训练，并引入卡尔曼滤波器实现在线实时预测。在多用户大规模多输入多输出（Multiple-Input Multiple-Output,MIMO）通信系统中，通过模拟分析了CSI参考信号周期与用户移动速度对系统性能的影响，并对比了不同预测方法的精度与复杂度。仿真结果表明，所提方法在30 km/h和60 km/h的用户速度下均显著优于传统方法，特别是在较大CSI参考信号周期下对小区边缘用户吞吐量的提升效果更为显著。 |
| 关键词: 大规模MIMO 时变信道预测 深度学习 门控循环单元 全卷积神经网络 |
| DOI：10.20079/j.issn.1001-893x.240805001 |
|
| 基金项目:国家自然科学基金资助项目(61701062)；重庆市基础与前沿研究计划项目(cstc2019jcyj-msxmX0079) |
|
| Massive MIMO Channel Prediction Method Based on Deep Learning |
| ZHOU Weia,b,XIANG Cenga,b,LIU Shixiao,CHEN Lia |
| (1a.School of Communication and Information Engineering;1b.Chongqing Key Laboratory of Mobile Communications Technology,Chongqing University of Posts and Telecommunications,Chongqing 400065,China;2.VIVO Mobile Communication Co.,Ltd.,Shenzhen 518049,China) |
| Abstract: |
| The authors propose a channel state information(CSI) prediction method that combines a time attention mechanism with an improved 2D fully convolutional neural network - gated recurrent unit(2DFCNN-GRU) model.The method extracts time-varying channel features through a feature pyramid network(FPN) and integrates feature fusion using a U-Net architecture.The GRU is employed for offline training,and a Kalman filter is introduced to achieve online real-time prediction.In a multi-user massive multiple-input multiple-output(MIMO) communication system,the impact of CSI reference signal period and user mobility on system performance is analyzed through simulations,and the accuracy and complexity are compared between different prediction methods.The simulation results show that the proposed method significantly outperforms traditional methods at user speeds of 30 km/h and 60 km/h,with a more notable improvement in cell-edge user throughput at larger CSI reference signal periods. |
| Key words: massive MIMO time-varying channel prediction deep learning gated recurrent unit fully convolutional neural network |
《电讯技术》编辑部 