An Efficient Attention Mechanism Based Neural Network for CSI Feedback in Massive MIMO System

In Frequency Division Duplex (FDD) systems, the quality and efficiency of Channel State Information (CSI) feedback enhancement are challenged by the large size of the CSI matrix in massive Multiple-Input Multiple-Output (MIMO) systems. A variety of machine learning techniques have been proposed to improve the accuracy of CSI feedback. However, most of these works achieved accuracy depending on the high computational complexity. To improve this situation, an efficient attention mechanism based neural network (EANet) for CSI feed-back is proposed in this paper. EANet reduces the computational complexity by simplifying the encoder and decoder network structure in the meantime it maintains the capabilities of feature extraction. Furthermore, a lightweight Multilayer Perceptron (MLP) module is integrated to improve the accuracy of CSI feedback by enhancing the ability to capture the channel features. Finally, experimental results demonstrate that EANet achieves higher feedback accuracy with a low computational complexity.