Untrained Neural Network based Symbol Detection for OTFS Systems

Abstract

Future mobile systems will support various high-mobility scenarios (e.g., unmanned aerial vehicles and high-speed trains) with strict mobility requirements. However, the current orthogonal frequency division multiplexing (OFDM) is unsuitable for these scenarios due to high inter-carrier interference (ICI) caused by high-mobility reflectors. Orthogonal time frequency space (OTFS) was proposed to overcome this challenge. Various state-of-the-art OTFS detectors have been investigated in the literature, including classical and training-based deep neural network (DNN) detectors. Classical OTFS detectors typically utilize matrix inversion operations, resulting in high computational complexity. Training-based DNN OTFS detectors outperform classical detectors regarding symbol error rate (SER) performance. However, these detectors rely on enormous computation resources and the fidelity of datasets for the training phase, and both are expensive. Deep image prior (DIP) has recently been proposed as an untrained DNN denoiser without training datasets. In this thesis, we develop a new real-time, decoder-only DNN architecture for DIP called D-DIP denoiser. We then combine the D-DIP denoiser with Bayesian-based parallel interference cancellation (BPIC) for symbol detection, resulting in the D-DIP-BPIC OTFS detector. Our simulation results demonstrate that the proposed D-DIP-BPIC OTFS detector outperforms state-of-the-art OTFS detectors regarding SER performance and computational complexity. To further enhance the performance of the D-DIP denoiser, we introduce a novel approach by incorporating a graph representation of wireless interference knowledge into the D-DIP denoiser, leading to the GDIP denoiser. Simulation results demonstrate that the proposed GDIP denoiser requires significantly fewer iterations than the D-DIP one to denoise the received signal. The combination of GDIP and BPIC shows an excellent SER performance under various OTFS configurations.