Abstract
Accurate channel estimation is essential for fully exploiting the advantages of orthogonal frequency division multiplexing (OFDM) systems. Compared with conventional estimators, deep neural networks can achieve higher estimation accuracy. Nevertheless, most existing deep learning–based approaches are purely data-driven, which limits their ability to effectively capture the inherent channel correlations. As a result, their performance degrades in dynamic environments and their interpretability remains limited. To address these challenges, this paper proposes a novel architecture, referred to as the Delay–Doppler Attention Network (D2AN). In particular, we design an attention mechanism where the attention coefficients are generated from basis functions sampled in the delay and Doppler domains. By incorporating the structure of channel correlation functions into the attention layers, D2AN explicitly models the underlying spectral and temporal dependencies of wireless channels. Simulation results demonstrate that the proposed D2AN significantly outperforms state-of-the-art deep learning–based counterparts as well as conventional channel estimators. Moreover, D2AN exhibits strong robustness across diverse signal-to-noise ratios, user mobilities, and delay spreads, while maintaining low computational complexity.
Funding Information
  • National Science and Technology Major Project (GXB-2-2025-5)

This publication has 11 references indexed in Scilit: