Joint Waveform and Beamforming Design in RIS-ISAC Systems: A Model-Driven Learning Approach

TL;DR

This paper introduces a model-driven learning approach for joint waveform and beamforming design in RIS-ISAC systems, enhancing radar sensing and communication performance while reducing computational complexity.

Contribution

It proposes a novel unfolding-based learning method for joint optimization of waveform and beamforming in RIS-ISAC, addressing computational challenges of prior techniques.

Findings

Achieves improved radar SINR and DoA estimation accuracy.

Reduces computational complexity compared to traditional methods.

Ensures communication QoS while optimizing sensing performance.

Abstract

Integrated Sensing and Communication (ISAC) has emerged as a key enabler for future wireless systems. The recently developed symbol-level precoding (SLP) technique holds significant potential for ISAC waveform design, as it leverages both temporal and spatial degrees of freedom (DoFs) to enhance multi-user communication and radar sensing capabilities. Concurrently, reconfigurable intelligent surfaces (RIS) offer additional controllable propagation paths, further amplifying interest in their application. However, previous studies have encountered substantial computational challenges due to the complexity of jointly designing SLP-based waveforms and RIS passive beamforming. In this paper, we propose a novel model-driven learning approach that jointly optimizes waveform and beamforming by unfolding the iterative alternative direction method of multipliers (ADMM) algorithm. Two joint design algorithms are developed for radar target detection and direction-of-arrival (DoA) estimation tasks in a cluttered RIS-ISAC system. While ensuring the communication quality-of-service (QoS) requirements, our objectives are: 1) to maximize the radar output signal-to-interference-plus-noise ratio (SINR) for target detection, and 2) to minimize the Cram\'{e}r-Rao bound (CRB) for DoA estimation. Simulation results verify that our proposed model-driven learning algorithms achieve satisfactory communication and sensing performance, while also offering a substantial reduction in computational complexity, as reflected by the average execution time.