Attention Guided Deep Learning for CSI Compression and Feedback

Accurate channel state information (CSI) feedback plays a vital role in improving the performance gain of massive multiple input multiple output (m-MIMO) systems. However, in obtaining accurate CSI, we usually encounter the problem that the overhead of CSI is too large, and the feedback bandwidth is limited. To solve the above issues, and considering that the channel estimation is not ideal, we propose a new deep neural network structure to complete the feedback of CSI, which is called Dilated Attention Inception Net (DAINet). Specifically, the network uses a variety of effective neural network structures including dilated convolution, attention mechanism, residual learning and Inception structure. DAINet extracts noise-free features from noisy CSI samples obtained from imperfect channel estimation to realize effective compression of CSI feedback. In the cases of ideal and imperfect channel estimation, considering the accuracy and training cost, the proposed DAINet performs better than the existing deep learning methods. Specifically, in terms of NMSE, in the case of ideal channel estimation, the NMSE of DAINet is about 1/2 of AnciNet and 1/12 of CsiNet; in the case of imperfect channel estimation, the NMSE of DAINet is about 9/10 of that of AnciNet and 1/2 of that of CsiNet. The number of network parameters of DAINet is about 80% of that of AnciNet and 130% of that of CsiNet. DAINet's FLOPs is 53% of AnciNet's.