基于Transformer大规模MIMO的CSI反馈网络

CSI feedback network for massive MIMO based on Transformer

针对频分双工大规模多输入多输出系统中信道状态信息反馈方法反馈精度低的问题,提出了一种名为CDTransformer的基于Transformer和卷积分解(convolutional decomposition,CD)的CSI反馈网络。CDTransformer将卷积分解融入改进的Transformer结构网络中,在不增加计算复杂度的情况下提升网络性能,并通过对网络的全连接层进行二值化实现轻量级部署。针对用户端功率有限的情况,提出了一种MixedTransformer网络模型。其中编码器采用计算成本较低、结构简单的单层卷积神经网络,而解码器则采用与CDTransformer模型相同的结构。CDTransformer模型融合了Transformer结构和卷积分解,提高了CSI反馈精度并实现轻量级部署。此外,引入了MixedTransformer模型,结合了CDTransformer和卷积神经网络的优点,以在功率有限情况下提供更好的性能。结果显示,相比于CsiTransformer网络模型,CDTransformer网络模型在归一化均方误差和余弦相似度方面分别提高了37.7%和0.2%。

In order to solve the problem of low feedback accuracy of channel state information feedback method in frequency-division duplex massive multi-input multi-output system,a CSI feedback network named CDTransformer based on Transformer and convolutional decomposition(CD)is proposed.CDTransformer incorporates convolutional decomposition into an improved Transformer network architecture to improve network performance without adding computational complexity,and enables lightweight deployment by binarizing the full connectivity layer of the network.A MixedTransformer network model is proposed for the limited power of the client.The encoder adopts a single layer convolutional neural network with low computational cost and simple structure,while the decoder adopts the same structure as the CDTransformer model.The CDTransformer model incorporates the Transformer structure and convolutional decomposition to improve CSI feedback accuracy and enable lightweight deployment.In addition,the MixedTransformer model was introduced,combining the advantages of CDTransformer and convolutional neural networks to provide better performance in power limited situations.The results show that compared with the CsiTransformer network model,the normalized mean square error and cosine similarity of CDTransformer network model are improved by 37.7%and 0.2%,respectively.