JCANet: Multi-domain federated lightweight self-attention CSI feedback network

Abstract

In frequency division duplex (FDD) massive MIMO systems, as the number of antennas increases, the amount of downlink channel state information (CSI) data fed back from the user’s end increases significantly, many deep learning (DL)-based CSI compression feedback methods show their potential. Existing networks mostly extract channel features in the angle-delay domain through complex convolutional structures, neglecting the frequency correlation among subcarriers, which makes it difficult to fully capture global features with long-distance dependencies. Moreover, these approaches suffer from high complexity. To address these issues, we propose a multi-domain joint lightweight self-attention feedback network (JCANet). First, a multi-domain joint strategy is proposed at the encoder side. On the basis of designing angular-delay domain convolution to extract local features of channel information, a frequency domain convolution (FCv) branch is used to span multiple subcarriers to capture the global features of the channel, achieving multi-domain extraction of channel information features. Then, a lightweight multi-scale cross-layer self-attention (LMSCA) module is proposed on the decoder side, which utilizes the multi-scale information of the CSI matrix to establish correlations and long-range dependencies between input sequences under low complexity. Simulation results show that JCANet achieves higher performance with lower computational complexity compared to other lightweight networks.