Joint Channel, Data, and Radar Parameter Estimation for AFDM Systems in Doubly-Dispersive Channels

TL;DR

This paper introduces joint channel, data, and radar parameter estimation schemes for AFDM systems in doubly-dispersive channels, enabling integrated sensing and communications with improved performance over existing methods.

Contribution

It presents a unified Bayesian framework for JCDE and RPE applicable to multiple waveforms, including AFDM, OTFS, and OFDM, with novel algorithms optimized for doubly-dispersive channels.

Findings

JCDE in AFDM outperforms SotA with a single pilot.

AFDM-based RPE surpasses OTFS and SBL techniques.

Proposed schemes significantly improve estimation accuracy.

Abstract

We propose new schemes for joint channel and data estimation (JCDE) and radar parameter estimation (RPE) in doubly-dispersive channels, such that integrated sensing and communications (ISAC) is enabled by user equipment (UE) independently performing JCDE, and base stations (BSs) performing RPE. The contributed JCDE and RPE schemes are designed for waveforms known to perform well in doubly-dispersive channels, under a unified model that captures the features of either legacy orthogonal frequency division multiplexing (OFDM), state-of-the-art (SotA) orthogonal time frequency space (OTFS), and next-generation affine frequency division multiplexing (AFDM) systems. The proposed JCDE algorithm is based on a Bayesian parametric bilinear Gaussian belief propagation (PBiGaBP) framework first proposed for OTFS and here shown to apply to all aforementioned waveforms, while the RPE scheme is based on a new probabilistic data association (PDA) approach incorporating a Bernoulli-Gaussian denoising, optimized via expectation maximization (EM). Simulation results demonstrate that JCDE in AFDM systems utilizing a single pilot per block significantly outperforms the SotA alternative even if the latter is granted a substantial power advantage. Similarly, the AFDM-based RPE scheme is found to outperform the OTFS-based approach, as well as the sparse Bayesian learning (SBL) technique, regardless of the waveform used.