Attention-Enhanced Channel Estimation for 6G MIMO in Unifying Far-Field and Near-Field

Given the challenges posed by the incompatibility of channel estimation (CE) across far-field and near-field scenarios, we propose a deep learning (DL)-based CE algorithm, i.e., the wavenumber-domain attention-enhanced neural network (WDANet). This method frames the CE problem as an image-denoising problem, enabling the recovery of channel state information (CSI) from received signals by leveraging unified channel representation in the wavenumber-domain (WD). Specifically, a wavenumber-embedding pre-processing step is designed to capture the clustered sparsity of propagating waves, forming an image-like structure by incorporating evanescent waves. To further enhance estimation accuracy, we introduce a multi-head channel attention module (MCAM) to improve noise extraction on the propagating wave components while limiting noise propagation to the evanescent wave components. Numerical results demonstrate that WDANet not only outperforms benchmark schemes but also exhibits robust generalization across diverse propagation environments.