Machine-Learning-Aided Optical OFDM for Intensity Modulated Direct Detection

End-to-end learning systems are conceived for Orthogonal Frequency Division Multiplexing (OFDM)-aided optical Intensity Modulation paired with Direct Detection (IM/DD) communications relying on the Autoencoder (AE) architecture in deep learning. We first propose an AE-aided Layered ACO-OFDM (LACO-OF...

Full description

Saved in:
Bibliographic Details
Main Authors: Xiaoyu, Zhang, Thien, Van Luong, Periklis, Petropoulos, Lajos, Hanzo
Format: Bài trích
Language:English
Published: IEEE 2022
Subjects:
Online Access:https://ieeexplore.ieee.org/document/9674226
https://dlib.phenikaa-uni.edu.vn/handle/PNK/5741
https://doi.org/10.1109/JLT.2022.3141222
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:End-to-end learning systems are conceived for Orthogonal Frequency Division Multiplexing (OFDM)-aided optical Intensity Modulation paired with Direct Detection (IM/DD) communications relying on the Autoencoder (AE) architecture in deep learning. We first propose an AE-aided Layered ACO-OFDM (LACO-OFDM) scheme, termed as LACONet, for exploiting the increased bandwidth efficiency of LACO-OFDM. LACONet employs a Neural Network (NN) at the transmitter for bit-to-symbol mapping, and another NN at the receiver for recovering the data bits, which together form an AE and can be trained in an end-to-end manner for simultaneously minimizing both the BER and PAPR. Moreover, the detection architecture of LACONet is drastically simplified compared to classical LACO-OFDM, since the Fast Fourier Transform (FFT) operation is applied only once at the receiver. We further propose a generalized AE-aided optical OFDM scheme for IM/DD communications, termed as IMDD-OFDMNet, where the unipolarity of the Time Domain (TD) signal is no longer guaranteed by the Hermitian Symmetry, but rather by taking the absolute square value of the complex TD signal. As such, all available subcarriers of IMDD-OFDMNet are used for carrying useful information, hence it has a higher throughput than the LACO-based schemes. As a further benefit, its transceiver requires only a single Inverse FFT or FFT. Finally, simulation results are provided to show that our learning schemes achieve better BER and PAPR performance than their conventional counterparts.