Learning Linear Filters for Underwater OFDM Channel Estimation with Attention-Aided MMSE

This letter presents an Attention-aided MMSE (AMMSE) channel estimation method for underwater OFDM systems. To address the severe time variation and multipath effects in underwater acoustic channels, AMMSE leverages a Transformer to learn a linear MMSE filter from data, capturing temporal and spectral correlations. Inference involves only a matrix-vector multiplication, ensuring low complexity. Simulations show that AMMSE outperforms LS, 1D-MMSE, and MMSE across all SNRs, with significant gains in low-SNR conditions.