BFP-Net: A DL-based ISAC beamforming prediction method for extended vehicle

Abstract

To enable the next generation of connected autonomous vehicles, the millimeter wave (mmWave)-based integrated sensing and communication (ISAC) system will be a critical technology in future vehicle-to-everything (V2X) networks. However, the rapid mobility of vehicles and the narrow beamwidth of mmWave signals present significant challenges for beam alignment, and point-target modeling methods often lead to substantial overhead, high latency, and complications. To address these issues, in this paper, a hybrid analog-digital (HAD) multi-input multi-output (MIMO) ISAC framework is adopted for the mmWave-based V2X network to reduce hardware costs and power consumption. Then, considering the narrow beamwidth of the mmWave system, the vehicle is modeled as an extended surface target with multiple scattering points, and a new association technique for these points is developed to improve prediction accuracy. Hence, a deep learning (DL)-based beamforming prediction network, namely beamforming prediction network (BFP-Net), is designed according to the ISAC signal beam prediction protocol and enables roadside units (RSUs) to transmit ISAC signals effectively for both downlink communication and sensing operations. The BFP-Net leverages a convolutional neural network long-short-term memory (CNN-LSTM) architecture to capture spatial and temporal correlations, providing enhanced modeling capabilities for beam prediction. Moreover, for highly dynamic vehicles, the BFP-Net predicts optimal beams for future time slots by extracting features from the received echo signals and eliminates the repetitive beam training inherent in the traditional communication protocol. Simulation results demonstrate that the proposed method significantly outperforms extended Kalman filter (EKF)-based methods in the mmWave V2X scenario, achieving higher beam gains and better performance for high-speed vehicles, and substantially reduces the overhead associated with beam training compared to the conventional neural network relying on pilot signals.