基于两步噪声学习网络的波束域毫米波大规模MIMO信道估计

Two-step Noise Learning Network Based Channel Estimation for Beamspace mmWave Massive MIMO

针对波束域毫米波大规模多输入多输出(Multiple-Input Multiple-Output,MIMO)系统,构建了一种新型两步噪声学习网络(Two-step Noise Learning Network,TNLNet)。基本原理是在接收信号反复经过卷积层和池化层提取噪声特征的基础上,利用波束域毫米波大规模MIMO信道矩阵稀疏性所引起的相邻元素相近的特点,采用下采样将信道矩阵重构成4个子矩阵,提高训练测试效率。该算法具有以比全卷积去噪近似消息传递(Fully Convolutional Denoising Approximate Message Passing,FCDAMP)算法和学习去噪的近似消息传递(Learned Denoising-based Approximate Message Passing,LDAMP)算法更低的复杂度,取得了比最小二乘算法、最小均方误差算法、FCDAMP和LDAMP更优的归一化均方误差(Normalized Mean Squared Error,NMSE)性能;与快速灵活去噪卷积神经网络(Fast and Flexible Denoising convolutional neural Network,FFDNet)相比虽然复杂度略高,但具有更优的NMSE性能,且在单一训练模型中获得了比FFDNet更宽的信噪比适用范围,增强了实用性。

For beamspace millimeter-wave(mmWave)massive multiple-input and multiple-output(MIMO)system,a new two-step noise learning network(TNLNet)is proposed.Firstly,the noise characteristic is extracted from the received signals through the convolution and the pooling.Then,by utilizing similar characteristics of adjacent elements caused by sparse characteristics of beamspace mmWave massive MIMO,down-sampling is implemented.Finally,four sub-matrices are reconstructed in the channel matrix,so as to improve the training and the testing efficiency.The results show that TNLNet achieves better normalized mean squared error(NMSE)performance than Least Square,Minimum Mean Square Error,Fully Convolutional Denoising Approximate Message Passing(FCDAMP)and Learned Denoising-based Approximate Message Passing(LDAMP),with lower complexity compared with FCDAMP and LDAMP.Specially,although the complexity of TNLNet is slightly higher than that of Fast and Flexible Denoising Convolutional Neural Network(FFDNet),TNLNet has better NMSE performance.Especially,TNLNet is more practical than FFDNet in a single training model.