I developed a deep learning framework that performs beamforming prediction and user localization from high-dimensional CSI in mmWave MIMO systems. To reduce computational overhead, I introduced an adaptive triangular filter with attention-based weighting that compresses CSI while keeping important spatial-frequency features.

Highlights:

• Multitask learning for joint beam prediction and coordinate localization.
• Adaptive filter module reduces model parameters by up to 88% while preserving performance.
• Attention-based weighting enhances important CSI features during compression.
• Achieved 88.0% top-1 beamforming accuracy and 2.30 m localization error.
• Demonstrated strong robustness under low-SNR conditions through noisy data augmentation.

Impact:

This method provides a lightweight and efficient solution for real-world mmWave massive MIMO deployment, reducing computation and memory usage while maintaining high accuracy.

Paper: View Paper