ISAC Receiver Design: A Learning-Based Two-Stage Joint Data-and-Target Parameter Estimation

This letter proposes a deep neural network Transformer-based sliding symbol detection for integrated sensing and communications (ISAC) in Single-Input-Multiple-Output Orthogonal Frequency Division Multiplexing (SIMO-OFDM) systems. The ISAC receiver design aims to both detect communication data symbols and estimate the target parameters simultaneously. Accordingly, we design a learning-based receiver and we prove that a sliding Transformer (STransformer) offers robust performance in signal detection tasks with much less training required compared with the conventional Sliding Bidirectional Recurrent Neural Network (SBRNN). For radar channel parameter estimation, both angle of arrival (AOA) and time delay (TD) are estimated through a two-step estimation. Analysis for the two parameters is based on the channel matrix estimated through Least Square (LS) according to the symbols previously detected. Our simulations show that the model is robust to various channel conditions tested and all trainings are done from scratch with only one epoch required. We further show that Transformer outperforms the conventional RNN-based model with limited training and has the potential to replace RNN models with better performance.