Learning-Based Constellation Design for Uplink Bi-Static Integrated Sensing and Communication

Abstract

This paper proposes an end-to-end deep learning based constellation design for integrated sensing and communication (ISAC) for the uplink of orthogonal frequency division multiplexing (OFDM) systems. Utilizing an auto-encoder architecture, the system designs and optimizes constellation mappings to balance the trade-off between communication and sensing performance under a bi-static scenario where receiver has no knowledge about transmitted signals. The constellation design is trained to adapt to specific channel conditions, offering flexible control over the communication-sensing performances by adjusting a radar weighting factor. Simulation results show that this design outperforms conventional constellation formats such as 64-QAM and 64-PSK in symbol error rate (SER), while outperforming the 64-QAM in sensing error. Furthermore, the proposed constellation design demonstrates robust performance even under channel state information (CSI) errors of up to 1.5%.