Neural-Network-Based Nonlinear Tomlinson-Harashima Precoding for Bandwidth-Limited Underwater Visible Light Communication.

Underwater visible light communication (UVLC) based on light-emitting diodes (LEDs) is considered a potential candidate for underwater wireless data transmission. However, the implementation of high-speed UVLC remains a challenge due to bandwidth limitation and nonlinear effects. In this paper, we investigate the performance of Tomlinson-Harashima precoding (THP) in the bandwidth-limited UVLC system for the first time. Apart from research on linear and Volterra series-based nonlinear THP, neural network (NN)-based THP is first proposed and conducted. At the transmitter, the intersymbol interference (ISI) and nonlinearity can be partially mitigated through a feedback neural network (FBN) without error propagation. A modulo operator is employed to eliminate instability. In addition, for the precoded signals, the power spectral density (PSD) is near-white. Therefore, it is possible to avoid the spectrum-shaping-induced signal-to-noise ratio (SNR) loss problem that is common in pre-emphasis methods. At the receiver, another modulo operator is used for decoding. An adaptive feedforward neural network (FFN) is applied to compensate for the channel state information (CSI) mismatch of the FBN and mitigate the remaining ISI and nonlinear impairments. In the demonstrated UVLC system, the experimental results prove the feasibility of the proposed method. The Q factor is increased by 3.39 dB at a data rate of 2.2 Gbps. 630 MBd CAP-16 transmission under a 7% HD-FEC threshold is realized, which is 90 MBd higher than that when employing linear postequalization only. [ABSTRACT FROM AUTHOR]