Predictive Beamforming Using Feature Fusion on Incomplete Time-Series Multimodal Sensing Data

Next-generation communications demand high-speed data transmission, making millimeter-wave (mmWave) and terahertz (THz) technologies essential. Reducing beam-forming overhead has led to the integration of sensing-assisted predictive beamforming with deep learning. However, in high-mobility scenarios, real-time collection of complete time-series data is often challenging due to network instability and device limitations. Motivated by this, we propose an LSTM-based method for completing incomplete time-series sensing data and develop a multimodal sensing-assisted predictive beamforming model. Experimental results show that the position data imputation error remains within meters, and training with the completed data improves model performance, achieving over 50% Top-1 and 93.53% Top-5 accuracy. This approach enhances prediction accuracy while reducing beamforming overhead