SRTRL: Shared Radar-Communication Temporal Representation Learning for Beam Prediction in Vehicular and UAV mmWave Networks
Abstract
Proactive beam management is essential for high-mobility millimeter-wave communication systems, where transmission latency may cause beams selected from current observations to become mismatched with the optimal beams at future transmission times. This paper investigates radar-only causal multi-horizon proactive beam prediction, aiming to directly rank future communication beams from past radar observations. To this end, we propose Shared-Beam-GCN-TCN, a multi-task framework that learns a shared predictive sensing-communication representation and performs structured decoding in the beam space. Specifically, four consecutive range-angle radar maps are encoded by a frame-wise convolutional neural network and a causal temporal convolutional network to capture short-term motion dynamics. The shared temporal representation is jointly optimized for future beam prediction and short-term radar prediction, encouraging the extraction of predictive sensing dynamics relevant to future communication states. A beam-space graph convolutional decoder then exploits the adjacency structure among 64 beam classes to model correlations between neighboring beams and improve future beam ranking. The framework is evaluated in two high-mobility mmWave communication scenarios, including a trajectory-disjoint DeepSense 6G vehicular scenario and a simulated unmanned aerial vehicle (UAV) scenario. In the vehicular scenario, Shared-Beam-GCN-TCN achieves the best overall performance across different prediction horizons, improving the average Top-3 and Top-5 accuracies by 8.2 and 6.9 percentage points over Seq-CNN-TCN, while reducing the mean beam-index error. In the UAV scenario, the proposed framework achieves an accuracy of 93.4% and a beam-index error of 1.022. These results demonstrate the effectiveness of shared predictive representation learning and structured beam-space decoding for causal multi-horizon beam prediction in high-mobility environments.
