Deep CSI Feedback for FDD MIMO Systems

With the increasing number of antennas at the base station (BS), the feedback overhead of traditional codebook in frequency division duplexing (FDD) mode becomes overwhelming, since the number of codewords in codebook increases quickly. Alternatively, we can directly feedback the channel state information (CSI) to the BS for precoding. To reduce the overhead of CSI feedback, this paper proposes three CSI compression models based on autoencoder network. The first two of them, adopting deep learning (DL) structure, are named FCNet and CNet, respectively. FCNet employs full-connected network architecture, while CNet is designed based on convolutional neural network with lightweight convolution kernels and multi-channel architecture. By applying principal component analysis (PCA) on CSI feedback, the third one, i.e., PCANet, is also studied and analyzed in details. Experiments show that CNet has best accuracy performance at the cost of high computational complexity, while FCNet shows medium accuracy and complexity among the three models. Besides, the accuracy of PCANet is nearly the same as CNet in some specific channel conditions. Compared with the state-of-the-art of CsiNet, the proposed models have their own advantages and limitations in different scenarios.