CHAST: Attention Aided SISO OFDM Channel Estimation

Next-generation wireless systems demand channel estimators that adapt to diverse propagation environments without requiring explicit channel statistics. We propose CHAST (Channel Attention eSTimation), a lightweight deep learning architecture for OFDM channel estimation that operates directly on sparse pilot observations at DM-RS positions. CHAST combines a CNN feature extractor with a single multi-head self-attention block to capture both local and long-range time-frequency dependencies. Unlike existing transformer-based approaches, CHAST requires no knowledge of channel governing parameters (SNR, Doppler shift, delay spread) while achieving comparable performance with significantly reduced complexity. Extensive evaluation on 3GPP channel models demonstrates that CHAST outperforms traditional methods, particularly in high-mobility scenarios where Kronecker covariance assumptions break down. Attention visualization reveals that the model learns physically meaningful estimation strategies, with attention neighborhoods dynamically expanding as SNR increases.