AI-Powered Multi-Modal Beam Prediction: A Combined Approach by Transfer Learning and Feature Selection

Extensive beam training overhead for optimal beam alignment is a remarkable challenge that makes the deployment of millimeter wave (mmWave) and terahertz (THz) communication systems with large antenna arrays hard. That said, sophisticated methods may replace the conventional beam training overhead. To this end, we propose a novel machine learning-based approach that leverages multi-modal data, combining positional and visual information from the wireless communication environment for efficient beam prediction. Our methodology implements a three-stage framework: transfer learning for initial feature extraction, and it is combined by feature engineering for positional information and systematic feature selection for vision information to identify optimal characteristics, and machine learning model to spot optimal beam. We compared our proposed approach with the original study used the same dataset and scenarios, and experimental results demonstrated significant improvements over the existing methodology used in majority of the results. It achieves the maximum gains for only position in Scenario 5 throughout the top-k accuracies. The gain for top-1 accuracy 35 %, and it is 17 % and 10 % for top-2 and top-3 accuracies in order. The proposed approach also reduces training time by using more simple machine learning algorithms. These findings illustrate that our multi-modal data-driven approach effectively addresses beam alignment challenges in next-generation wireless communications by reducing the beam training overhead and achieve high performance simultaneously, enabling more efficient and reliable mmWave and THz communication systems.