Predictive Beamforming for Vehicles With Complex Behaviors in ISAC Systems: A Deep Learning Approach

Integrated sensing and communication (ISAC) technology enables reliable communication and highly accurate sensing performance simultaneously in vehicular networks. In the traditional paradigm, an effective predictive beamforming scheme requires high computational complexity caused by the need of keep tracking vehicle motion parameters. To address this challenge, in this paper, we propose a deep learning (DL)-based framework that bypasses explicit channel prediction, thereby reducing the signaling overhead and facilitating the direct optimization of the beamformer to enhance the performance of the ISAC system. In particular, we aim to address the limitations of conventional ISAC beam-tracking techniques, which are inadequate for on-road vehicles with intricate behaviors. Specifically, we design an encoder-prediction-decoder network (EPDnet) to optimize a network utility function based on the joint Cramér-Rao Bounds (CRBs) for evaluating sensing performance while ensuring an acceptable downlink communication sum-rate. To further exploit the spatio-temporal correlations from the data, the proposed network adopts multiple convolutional neural network (CNN)-based encoders to capture historical channel features for each vehicle. Additionally, long short-term memory (LSTM) modules are employed to predict high-level features at the subsequent time slots. A fully connected network (FCN)-based decoder is then adopted to acquire the optimized beam. Simulation results demonstrate that the proposed EPDnet-based predictive beamforming design can achieve a favorable sensing performance while satisfying the requirement for reliable communications.