Parallel Computing for mmWave Beam Tracking Based on LSTM and Adversarial Sparse Transformer
Chaohui Xu, Liqun Fu · 2024
The 5G New Radio (5G NR) standard proposes beam management and beam tracking mechanisms, aiming to solve the blockage and beam misalignment in millimeter-wave (mmWave) communication. However, these usually become unrealistic due to the excessive time overhead. To address this issue and minimize the need for frequent beam training in high-speed scenarios, we aim to find a high accuracy and low complexity method for beam tracking. In particular, we propose a deep learning (DL)-based beam tracking algorithm, in which long short-term memory (LSTM) is cascaded with sparse transformer. LSTM extracts user equipment (UE) movement features from the received signal vector, and the sparse transformer predicts the optimal beam for future time slots in parallel. By incorporating adversarial training and a specifically designed loss function, our algorithm can be extended to handle more complex scenarios. The experimental results show that our algorithm could improve the top-1 accuracy (top-1 acc) by 3.65% when the velocity of UE is 20 m/s, and the model parameter numbers and the giga floating-point operations per second (GFLOPs) are only 16.3% and 11.98% of the baseline. Our scheme significantly reduces the model parameters while effectively improving the accuracy of beam tracking.