Predictive Beamforming in Integrated Sensing and Communication-Enabled Vehicular Networks

Abstract

Integrated sensing and communication (ISAC) has recently attracted significant research attention. This paper develops the deep learning-based predictive beamforming method for the ISAC-enabled vehicular networks. Traditional deep learning (DL) is a data-driven approach, which means that numerous training samples are required to improve system performance. In addition, embedded devices are not able to provide sufficient computing power, which hinders the application of DL solutions. Motivated by this, the dynamic self-attention mechanism is proposed to reduce the dependence of DL on training samples. Aiming for the optimal trade-off between sensing performance and computational complexity, the efficient model design, Self-Attention Channel Shuffle Mobile Network (SACSMN), is formulated. Experimental results demonstrate that SACSMN achieves similar sensing performance to that based on the full training set under the condition of few samples, the dependence of SACSMN on training samples is significantly reduced. Furthermore, SACSMN significantly reduces the computational complexity while achieving the same level of sensing performance as the benchmarks, realizing the optimal trade-off between system sensing performance and computational complexity. Benefiting from the robust sensing performance of SACSMN, the system achieves the same level of communication performance as that based on full training samples in the case of few samples.