Mixed Model- and Data-Driven Spatial Non-Stationary Channel Estimation for Near-Field XL-MIMO Systems

In this paper, we investigate the problem of spatial non-stationary channel estimation for near-field extremely large-scale multiple-input multiple-output (XL-MIMO) systems. Specifically, we propose a mixed model- and data-driven deep learning (DL) approach that achieves accurate near-field channel estimation under spatial non-stationary effects while maintaining low computational complexity. Our approach first exploits the polar-domain sparsity of the near-field channel to develop an unfolding network for efficient on-grid sparse channel recovery. This network is designed by unfolding the inverse-free variational Bayesian inference (IF-VBI) algorithm, referred to as VBI-Net. Moreover, to mitigate the energy leakage effect inherent in ongrid methods and the estimation error introduced by spatial non-stationary effects, we further enhance channel estimation accuracy by utilizing a data-driven deep learning framework. Simulation results validate that our proposed mixed model- and data-driven DL approach significantly outperforms existing state-of-the-art schemes, validating its effectiveness and superiority.