A Machine Learning Enabled Single-Photodetector Framework for 5G/6G Hybrid Digital/Analog Radio-Over-Fiber Links

The never ending persuit of higher data rate has led to the proposal and utilization of higher and higher radiofrequency (RF) for wireless communication systems, e.g. millimeter-wave for 5G and terahertz range for 6G. In the context of an analog radio-over-fiber (ARoF) link, high-speed photodiode is required to support detection and generation of such high RF signals. However, the bandwidth of the high speed photodiode may not be fully utilized due to the bandwidth restriction in the wireless domain. The available bandwidth of the high-speed photodiode may be repurposed for the detection of other signals, for example, to support base-band digital radio-over-fiber (DRoF) links or passive optical networks. However, the nature of photodiode heterodyne detection may be problematic, as the detection of ARoF signal, generates two components: one at RF for wireless transmission, and another at baseband which corresponds to the square amplitude of the ARoF signal. The presence of the baseband component leads to performance degradation of the baseband link. Research into mitigating the effect of square amplitude of the ARoF signal has been conducted. For example, in (A. V. Marti, N. Vokic, T. Zemen and B. Schrenk, 2022 Opt. Fiber Commun. Conf. and Exhib. (OFC), 1-3.), the DRoF signal is moved to an intermediate frequency, away from the baseband squared amplitude of the AROF signal. However, the DRoF signal may be required to be moved to a relatively higher frequency, as the bandwidth of the ARoF signal increases. On the other hand, the demonstration in (G. H. Thng and S. Mikki, in IEEE Access, 11, 114527–114535) proposed a simple signal processing approach to suppress the baseband impairment, by extracting the squared amplitude of the ARoF signal, through homodyning of the ARoF signal, and subsequently it is used to subtract from the baseband DRoF signal, to remove the squared amplitude of the ARoF component at baseband. However, as discussed in the paper, the squared amplitude of the ARoF signal extracted using the ARoF signal is not identical to the impairment at baseband. While the proposed simple signal processing approach may suppress the impairment caused by the ARoF signal, the residue impairment will still affect DRoF link performance, especially with increasing ARoF power. Hence, we explore the use of machine learning, tasked to predict the actual impairment at baseband by using the imperfectly extracted version, and subsequently regenerating the’ clean’ DRoF signal, to further improve the performance of a single photodiode framework supporting co-detection of hyrbid analog and digital RoF signals.