Learnable Sparse Transformation-Based Massive MIMO CSI Recovery Network

In frequency division duplex massive Multiple-Input Multiple-Output (MIMO) systems, plenty of Channel State Information (CSI) needs to be fed back. By exploiting the correlation of channel coefficients, the channel matrix can be transformed into a sparse form for compression. In this letter, we propose a model-based sparse recovery network that combines the advantages of compressed sensing reconstruction algorithms and neural networks, to perform CSI compression and reconstruction fast and accurately. Moreover, considering that the CSI is not strictly sparse in the discrete Fourier transform basis, we introduce a sparse autoencoder in our network to learn sparse transformations. Extensive experiments show that our model outperforms traditional compressed sensing algorithms and network-based methods.