85-GHz photonics W-band IM/DD PAM4 wireless transmission over 300 m based on nonlinear U-net symmetrical encoder-decoder equalizer

We propose a nonlinear U-Net equalizer based on a fully convolutional neural network for radio-over-fiber (ROF) communication systems. The U-Net encoder generates numerous feature channels through multi-layer downsampling, effectively extracting and transmitting deep temporal information. The symmet...

Full description

Saved in:
Bibliographic Details
Published in:Optics communications Vol. 577; p. 131423
Main Authors: Zhang, Jie, Zhou, Wen, Bian, Chengzhen, Ge, Jingtao, Xu, Sicong, Ma, Yuan, Wang, Qihang, Wang, Siqi, Ou, Zhihang, Hu, Sheng
Format: Journal Article
Language:English
Published: Elsevier B.V 01.03.2025
Subjects:
Encoder-decoder model
Photonics aided technology
U-net equalizer
W-band
Encoder-decoder model
Photonics aided technology
W-band
U-net equalizer
ISSN:0030-4018
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:We propose a nonlinear U-Net equalizer based on a fully convolutional neural network for radio-over-fiber (ROF) communication systems. The U-Net encoder generates numerous feature channels through multi-layer downsampling, effectively extracting and transmitting deep temporal information. The symmetric decoder side uses corresponding upsampling steps to restore resolution, ensuring that the output has the same size as the input, thus generating equalization for each symbol. To mitigate the spatial information loss caused by downsampling, U-Net introduces skip connections, ensuring that the upsampled feature maps retain more shallow temporal information. The symmetric structure facilitates a more balanced transmission of information, achieving equilibrium between global and local information, as well as deep and shallow features. By using our proposed U-Net equalizer, we experimentally demonstrated single-channel 4G baud 85 GHz PAM4 signal transmission over a 300-m wireless link, and its bit error rate (BER) is below the 7% hard-decision forward error correction (HD-FEC) threshold of 3.8 × 10−3. Experimental results show that compared to Volterra and other neural networks, the U-Net equalizer reduces the BER by 42%, reduces computational complexity by 57%, parameters by 13%, and training data size by 8%. To the best of our knowledge, this is the first time such a novel U-Net equalizer in ROF system has been proposed. Its high precision, low complexity, and reduced training data requirements make this equalizer highly promising for future millimeter-wave communication development and applications.
ISSN:0030-4018
DOI:10.1016/j.optcom.2024.131423