基于双重注意力机制的大规模MIMO系统CSI反馈

CSI feedback based on the dual attention mechanism for massive MIMO systems

在频分双工大规模多输入多输出(Multiple Input Multiple Output,MIMO)系统中,针对现有基于深度学习信道状态信息(Channel State Information,CSI)反馈方法重建精度低、未考虑量化损失以及压缩率固定的问题,提出一种基于注意力机制和密集连接卷积网络的多压缩率CSI反馈方案。该方法采用双重注意力机制提取CSI空间和时间相关性,串行全连接网络和密集连接卷积网络分别压缩和重构CSI。实验结果表明:在室内信道条件下,与基于卷积神经网络CsiNet相比,所提方法的参数量较少,归一化均方误差平均降低4.8 dB。所提反馈方法仅8比特量化可实现与未量化相近的重建精度。

In a frequency division duplex massive multiple⁃input multiple⁃output(MIMO)system,the existing deep learning based channel state information(CSI)feedback methods ignore the quantization loss,and output a low reconstruction accuracy and a fixed compression rate.In regard of this,this paper proposes a multiple compression ratio CSI feedback scheme based on the attention mechanism and the densely connected convolutional networks(DCCN).The method uses the dual attention mechanism to extract the spatial and temporal correlation of CSI,and adopts the serial fully connected layer and DCCN to achieve CSI compression and reconstruction.The results show that under indoor channel conditions,the proposed method use fewer parameters and its normalized mean square error is reduced by nearly 4.8 dB,compared with classic feedback methods.The proposed method can achieve the similar reconstruction accuracy as the unquantized does by using only 8⁃bit quantization.