Attention Mechanism for Beam Prediction in mmWave Communications with Mission-Critical Applications

Abstract

In this paper, we investigate beam prediction in a multi-antenna communication system serving mission-critical applications using millimeter-wave bands. To provide continuous service, the base station needs to sweep the beams in the codebook periodically. The signaling overhead for beam sweeping results in low resource utilization efficiency, and the feedback delay may lead to beam misalignment. To address these issues, we propose a beam prediction architecture by integrating an attention mechanism into a long-short-term memory (LSTM) network, with which the next beam direction and signal-to-noise ratio (SNR) are predicted based on different observations of channel state information. Specifically, we consider four types of observations: full observation, partial observation, binary observation, and single-beam observation. Our simulation results show that with the first three types of observations, the Euclidean distance between the predicted beam and the best beam is less than two with a probability of 95%. In addition, with partial observation, the prediction algorithm can achieve the best prediction accuracy in terms of SNR. Compared with full observation, by removing some irrelevant information, the attention mechanism can concentrate on the center of a cluster of beams, and achieve better performance with less feedback information. With partial or binary observations, the attention mechanism can reduce prediction errors by 50% compared to an existing LSTM network with the single-beam observation.