摘要
针对基于U-Net网络的信道估计会导致由于神经网络上下采样所造成的信息丢失的问题,文章提出了基于UNet++的一位大规模MIMO系统下的信道估计算法。该算法首先通过稠密的卷积块将深层特征与浅层特征相连接,大幅降低神经网络上下采样所造成的信息丢失,其次运用条件生成对抗网络(cGAN)对其进行训练,引入自适应损失函数以正确训练网络,并完成根据导频的量化接收信号来进行信道矩阵的估计。仿真结果表明,相同条件下,该信道估计方法与现有其他深度学习(DL)方法相比具有更好的表现,且在导频序列较短或信噪比(SNR)较低的情况下,具有很强的鲁棒性。
Aiming at the problem that channel estimation based on U-Net network will lead to information loss caused by neural network up-sampling and downsampling,this paper proposes a channel estimation algorithm under a one-bit massive MIMO system based on UNet++.Firstly,the algorithm connects the deep features with the shallow features through dense convolutional blocks to greatly reduce the information loss caused by the upper and lower sampling of the neural network,and secondly,it uses the conditional generative adversarial network(cGAN)to train it,introduces an adaptive loss function to train the network correctly,and completes the estimation of the channel matrix according to the quantized received signal of the pilot.The simulation results show that under the same conditions,the channel estimation method performs better than other existing deep learning(DL)methods,and has strong robustness under short pilot sequences or low signal-to-noise ratio(SNR).
作者
宁琦
许鹏
王韵乔
林嘉怡
NING Qi;XU Peng;WANG Yun-jiao;LIN Jia-yi(School of Electronic Information Engineering,Shenyang University of Aeronautics and Astronautics,Shenyang 110136,China)
出处
《电脑与信息技术》
2024年第1期73-77,共5页
Computer and Information Technology
基金
辽宁省教育厅系列项目(项目编号:LJKMZ20220535)
辽宁省自然科学基金(项目编号:2022-MS-294)
沈阳航空航天大学博士科研启动项目(项目编号:19YB03)。
