Frequency Division Sparse Neural Network for CSI Feedback in Massive MIMO Systems

In massive multiple-input and multiple-output (mMIMO) systems, it is critical to acquire accurate channel state information (CSI) to guarantee the efficiency and robustness of wireless communications. However, sending the CSI to the transmitters requires extra bandwidth consumption with transmission delays, especially for the frequency division duplexing (FDD)scheme. According to our research, a CSI matrix has different information densities in different frequency domains. Thus, in this paper we propose a frequency division processing deep learning framework, namely Frequency Division Sparse Neural Network (FDSNet), to compress and recover CSI with very low bandwidth consumption in this paper. The framework processes and fuses CSI features from different frequencies, enhancing the compression and recovery ability of the network. Our experimental results demonstrate that the matrix recovery accuracy of the proposed FDSNet outperforms the existing methods based on deeplearning under the condition that the computational complexity and the amount of parameters are close.