Neural Network-based Single-carrier Joint Communication and Sensing: Loss Design, Constellation Shaping and Precoding

We investigate the impact of higher-order modulation formats on the sensing performance of single-carrier joint communication and sensing (JCAS) systems. Several separate components such as a beamformer, a modulator, a target detector, an angle of arrival (AoA) estimator and a communication demapper are implemented as trainable neural networks (NNs). We compare geometrically shaped modulation formats to a classical quadrature amplitude modulation (QAM) scheme. We assess the influence of multi-snapshot sensing and varying signal-to-noise ratio (SNR) on the overall performance of the autoencoder (AE)-based system. To improve the training behavior of the system, we decouple the loss functions from the respective SNR values and the number of sensing snapshots, using upper bounds of the sensing and communication performance. The NN-based sensing outperforms classical algorithms, such as a Neyman-Pearson based power detector for object detection and ESPRIT for AoA estimation for both the trained constellations and QAM at low SNRs. We show that the gap in sensing performance between classical and shaped modulation formats can be significantly reduced through multi-snapshot sensing. Lastly, we demonstrate system extension to multi-user multiple-input multiple-output (MIMO) to address the improvement of spatial efficiency when servicing multiple user equipments (UEs). Our contribution emphasizes the importance of estimation bounds for training neural networks, especially when the trained solutions are deployed in varying SNR conditions.