Deep Learning Based Phase Noise Tolerant Radio-Over-Fiber Receiver

In recent years, incoherent approaches in the generation, transport, and detection of millimeter-wave Radio-over-Fiber signals have attracted a lot of attention due to their inherent technological simplicity and cost-effectiveness, which is however at the expense of additional phase-induced noises caused at the receiver's output. The power of deep learning, a subset of machine learning, has appeared recently to be very effective to improve the performance of communication blocks, particularly in signal compression, signal detection, and end-to-end communications. In this article, we propose and demonstrate a new receiver architecture by incorporating deep learning at the receiver. The proposed receiver is demonstrated on an unlocked heterodyning Radio-over-Fiber link. Results show that the proposed deep learning based receiver exhibits a greater tolerance against phase-induced noises, with a bit error rate improvement from
$10^{-1}$
to
$10^{-5}$
. In addition, the proposed deep learning based receiver performs better, in terms of bit error rate, than conventional self-homodyning based approach when the frequency spacing between reference tone and the main data signal is small.