Near-Field Channel Estimation and Sparse Reconstruction for FDD XL-MIMO Systems

The exponential growth of antennas in extremely large-scale MIMO (XL-MIMO) systems can lead to substantial overhead in pilot transmission for channel estimation and feedback, resulting in a decline in spectrum efficiency. This letter proposes a deep learning (DL)-based framework tailored for frequency division duplex (FDD) XL-MIMO, focusing on specialized neural networks for channel estimation and sparse reconstruction. For channel estimation, we design frequency-aware pilots by using dense layers according to the signal model and develop an attention mechanism-based residual channel estimation (A-RCE) network, which leverages inherent correlations within the channel matrix across subcarriers and antennas to improve estimation accuracy. To reduce channel state information (CSI) feedback overhead, we introduce a trainable fast iterative shrinkage thresholding (TFIST) network that leverages the polar-domain sparsity of the near-field channel to achieve a low-dimensional sparse representation. The simulation results validate the effectiveness of our proposed scheme, which can significantly enhance the estimation performance compared to other benchmark schemes.