Fusion of self-attention mechanism for CSI feedback in massive MIMO systems

Abstract

In massive MIMO systems, obtaining accurate channel state information (CSI) is crucial for optimal channel coding and beamforming. However, traditional CSI feedback methods require high bandwidth and also consume a large amount of power and computing resources. To address these challenges, several compressed sensing-based techniques have been implemented in recent years. These techniques, however, are often iterative and computationally complex to implement in power-constrained user equipment. In this paper, we propose a novel fusion of the self-attention mechanism, called FSAMNet, to efficiently and accurately implement the CSI feedback task for massive MIMO systems. Our proposed FSAMNet adopts both the residual connections in the attention mechanism and a sequence of depth-separable convolutional layers to enhance the model’s performance and expressive ability. Specifically, we apply a multi-layer self-attention mechanism in the encoder part to achieve feature extraction and compression. In the decoder part, we use multiple convolutional layers and self-attention mechanisms to convert the embedding vector generated by the encoder back into the original image. Experimental results show that the performance of our proposed FSAMNet outperforms conventional benchmark schemes in terms of feedback network performance.