| 摘要: |
| 针对空中大规模多输入多输出（Multiple-Input Multiple-Output，MIMO）系统中宽带多用户混合波束赋形设计面临的高计算开销问题，提出了一种基于深度学习的混合波束赋形网络（Attention Mechanism Based on Hybrid Beamforming Network，AMHBNet）框架。该框架通过端到端神经网络对时分双工和频分双工系统的关键传输模块进行一体化建模。对于时分双工系统，AMHBNet联合优化上行导频组合与下行混合波束赋形；对于频分双工系统，联合优化下行导频传输、上行信道状态信息反馈和下行混合波束赋形，避免了显式信道重建，减少了导频和反馈开销。最后，AMHBNet结合卷积神经网络（Convolutional Neural Network，CNN）与基于混合波束赋形的注意力机制（Attention Mechanism Based on Hybrid Beamforming，AMHB），利用CNN提取隐式信道特征并捕捉复杂非线性关系，同时通过AMHB机制中的随机特征映射、低维近似和高效计算策略，降低计算复杂度，保持高精度和稳定性。仿真结果表明，与基于CNN的混合波束赋形网络（CNN-based Hybrid Beamforming Network，CNN-HBFN）方案相比，AMHBNet在和速率上平均提升6.5%，计算复杂度平均降低5.8%。 |
| 关键词: 空中大规模MIMO系统 混合波束赋形 深度学习 |
| DOI：10.20079/j.issn.1001-893x.241009003 |
|
| 基金项目:国家自然科学基金资助项目（62371262） |
|
| Deep Learning-driven Hybrid Beamforming in Aerial Massive MIMO Systems |
| JI Honghui,SUN Qiang,LI Zejun,CHEN Xiaomin |
| (a.School of Artificial Intelligence and Computing;b.School of Information Science and Technology，Nantong University，Nantong 226019，China) |
| Abstract: |
| For the challenge of high computational overhead in wideband multi-user hybrid beamforming for aerial massive multiple-input multiple-output(MIMO) systems，a deep learning(DL)-based hybrid beamforming network framework，named the attention mechanism based on hybrid beamforming network(AMHBNet)，is proposed.The framework employs an end-to-end neural network to model key transmission modules for both time division duplex(TDD) and frequency division duplex(FDD) systems.For TDD systems，AMHBNet simultaneously optimizes uplink pilot combining and downlink hybrid beamforming.For FDD systems，AMHBNet jointly optimizes downlink pilot transmission，uplink channel state information(CSI) feedback，and downlink hybrid beamforming.This approach eliminates the need for explicit channel reconstruction while reducing pilot and feedback overhead.Finally，AMHBNet integrates convolutional neural networks(CNNs) with an attention mechanism for hybrid beamforming(AMHB).The CNNs extract implicit channel features and model complex nonlinear relationships.The AMHB mechanism uses random feature mapping，low-dimensional approximation，and efficient computation strategies to reduce complexity while maintaining accuracy and stability.Simulation results demonstrate that AMHBNet outperforms the CNN-HBFN scheme.It achieves an average improvement of 6.5% in sum rate and reduces computational complexity by an average of 5.8%. |
| Key words: aerial massive MIMO systems hybrid beamforming deep learning |
《电讯技术》编辑部 