SE-Attention Enhanced Sensing-Aided CSI Feedback for mMIMO Systems

Massive multiple input multiple output (mMIMO) systems require accurate channel state information (CSI) feedback but face high overhead. Existing deep learning methods ignore environmental information, while sensing-assisted frameworks reduce dimensionality at the cost of noise amplification in channel reconstruction. To address this, we enhance the RENet recovery network and propose RENet+, which incorporates a squeeze-and-excitation (SE) attention mechanism to adaptively recalibrate channel features. This design suppresses redundant components while emphasizing critical angular-spread information. As the first work integrating SE attention into sensing-aided CSI recovery, the proposed method significantly improves reconstruction accuracy under low feedback overhead. Evaluated on Saleh-Valenzuela channel models with 5 scatterers, the jointly trained JNet-Joint+ achieves NMSE gains of 4.7 dB and 6.4 dB over CsiNet at compression ratios (CR) of 64 and 128, respectively, and outperforms the original JNet-Joint by 0.9 dB (CR=64) and 0.7 dB (CR=128).