A Millimeter-Wave Beam Tracking Method Based on Deep Active Learning

In this paper, we address the challenge of Angle of Arrival (AoA) tracking in millimeter-wave (mmWave) communication systems, where fixed codebooks impose fundamental limitations on estimation accuracy. To overcome this, we propose a deep active learning-based adaptive beamforming scheme that integrates deep neural networks (DNN) with particle filter (PF). The proposed approach dynamically adjusts beamforming based on AoA probability distributions, effectively mitigating the drawbacks of conventional codebook-based methods. Specifically, a DNN-based beamformer, continuously refined using online AoA estimates, generates adaptive beamformers guided by prior AoA probabilities. These beamformers provide informative signal measurements for a PF-based AoA estimator, enabling high-precision tracking. Simulation results indicate that the proposed scheme achieves an 50% improvement in tracking accuracy compared to existing codebook-based methods.