Variational Autoencoder-Based Parameter Estimation in Beam-Space OFDM Integrated Sensing and Communication

Abstract

In this work, we propose a framework based on Deep Neural Networks (DNNs) for radar parameter estimation in an Integrated Sensing and Communication (ISAC) system em-ploying a realistic and hardware-efficient Hybrid Digital-Analog (HDA) architecture that uses Orthogonal Frequency Division Multiplexing (OFDM) digital modulation. This framework takes raw signals as input and utilizes a Variational Autoencoder (VAE) followed by a regression network to output the spatial extent and location of extended targets. Owing to the HDA setup, the co-located radar receiver uses multi-block measurements to perform parameter estimation. The proposed solution is motivated as a remedy for the increasing computational complexity associated with high-resolution extended target estimation in multi-carrier digital modulations such as OFDM. In addition, it is well known that off-grid delay-Doppler shifts which are present in the doubly-dispersive channels in the high mobility scenarios expected in ISAC applications, exhibit leakage effects that adversely affect the parameter estimation performance. Due to the data-centric nature of the proposed method, these effects can be learned by the network. We provide numerical results to showcase the effectiveness of the proposed framework for parameter estimation. 1