Deep Learning-Aided Optical IM/DD OFDM Approaches the Throughput of RF-OFDM

• Published in: IEEE journal on selected areas in communications Vol. 40; no. 1; pp. 212 - 226
• Main Authors: Van Luong, Thien, Zhang, Xiaoyu, Xiang, Luping, Hoang, Tiep M., Xu, Chao, Petropoulos, Periklis, Hanzo, Lajos
• Format: Journal Article
• Language: English
• Published: New York IEEE 01.01.2022; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Artificial neural networks; autoencoder (AE); bit error ratio (BER); Complexity; Complexity theory; Deep learning; deep neural network (DNN); Machine learning; O-OFDMNet; Optical distortion; Optical orthogonal frequency division multiplexing (O-OFDM); Optical receivers; Optical signal processing; Optical transmitters; Orthogonal Frequency Division Multiplexing; Peak to average power ratio; peak-to-average power ratio (PAPR); Radio frequency; Random noise; Real time operation; Signal processing; Time domain analysis; Transceivers; visible light communications (VLC)
• ISSN: 0733-8716; 1558-0008
• DOI: 10.1109/JSAC.2021.3126080

Deep learning-aided optical orthogonal frequency division multiplexing (O-OFDM) is proposed for intensity modulated direct detection transmissions, which is termed as O-OFDMNet. In particular, O-OFDMNet employs deep neural networks (DNNs) for converting a complex-valued signal into a non-negative signal in the time-domain at the transmitter and vice versa at the receiver. The associated frequency-domain signal processing remains the same as in conventional radio frequency (RF) OFDM. As a result, our scheme achieves the same spectral efficiency as the RF scheme, which has never been attained by the existing O-OFDM schemes, because they have relied on the Hermitian symmetry of the spectral-domain signal to guarantee that the time-domain signal becomes real-valued. We show that O-OFDMNet can be viewed as an autoencoder architecture, which can be trained in an end-to-end manner in order to simultaneously improve both the bit error ratio (BER) and the peak-to-average power ratio (PAPR) for transmission over both additive white Gaussian noise and frequency-selective channels. Furthermore, we intrinsically integrate a soft-decision aided channel decoder with our O-OFDMNet and investigate its coded performance relying on both convolutional and polar codes. The simulation results show that our scheme improves both the uncoded and coded BER as well as a reducing the PAPR compared to the benchmarks at the cost of a moderate additional DNN complexity. Furthermore, our scheme is capable of approaching the throughput of RF-OFDM, which is notably higher than that of conventional O-OFDM. Finally, our complexity analysis shows that O-OFDMNet is suitable for real-time operation.