Analysis of an adaptive orbital angular momentum shift keying decoder based on machine learning under oceanic turbulence channels

Oceanic turbulence tends to degrade the performance of underwater optical communication (UOC) systems based on orbital angular momentum (OAM) shift keying (SK). A decoder for the UOC-OAM-SK using convolutional neural networks (CNNs) is investigated. We simulate 8 kinds of superposition Laguerre-Gaussian (LG) beams as a trinary OAM-SK encoder; these beams propagate under simulated oceanic channels. The results show that in temperature-dominated situations, the decoders based on the CNN have a high accuracy (nearly 100%) under weak-to-moderate turbulence and have an accuracy greater than 93% under strong turbulence at a distance of 60 m. Under weak-to-moderate turbulence, the accuracies are higher than 95% within 80 m, and under strong turbulence, the accuracies are lower than 90% after 60 m propagation. The decoder with an incorporated CNN is insensitive to the balance parameter in most situations, except for those that are salinity dominated. Furthermore, the CNN trained with a database mixed with several levels of turbulence has a higher accuracy when accommodating an unknown level of turbulence than when trained with a single level of turbulence. This work is expected to aid in the future design of UOC-OAM-SK systems.