Deep Learning-Driven Joint Beamforming and Channel Tracking via Sparse Bayesian Filtering for Massive MIMO

The temporal variation of massive multiple-input multiple-output (MIMO) channels poses significant challenges to accurate channel estimation and effective beamforming design. To address this problem, we propose an end-to-end deep learning scheme for channel tracking with beamforming optimization.

The proposed scheme integrates a convolutional neural network (CNN)-based beamformer fed by the sparse channel estimates to align with the dominant directions, together with an adaptive channel estimator using the online variational sparse Bayesian filtering (OVSBF) algorithm, which recursively updates the channel information at each time slot by leveraging temporal correlations. Numerical results demonstrate that the proposed approach achieves superior beam alignment and significantly outperforms other beamforming benchmarks.