Experimental study of machine-learning-based orbital angular momentum shift keying decoders in optical underwater channels

We experimentally demonstrate the performance of 16-ary orbital angular momentum shift keying (OAM-SK) decoders based on convolutional neural networks (CNNs) in an underwater optical wireless communication (UOWC) system. We verify the decoding accuracy of the OAM-SK system in a 1-m tank filled with flowing salty water. The oceanic turbulence is emulated by the random phase screen method by a spatial light modulator (SLM). The results show that the decoders are robust to short-distance water fluctuations induced by the water flow, with an accuracy of more than 99% in clean water. In turbid water, the decoders require input images with a greater number of pixels and can reach an accuracy of more than 99% with a pixel number of 64 × 64. The decoders are tolerant to weak turbulence, with accuracies greater than 96.78% and 99.75% for input pixel numbers of 32 × 32 and 64 × 64 in 20-m temperature-induced channels with an equivalent temperature structure parameter of C_n ²=10⁻¹⁵K²⋅m^−2∕3. As the strength of the oceanic turbulence increases, the accuracy decreases to lower than 85.5%. In both turbid channels and turbulent channels, large OAM modes are more likely to be incorrectly recognized. Increasing the pixel number of the input images or decreasing the order of the coding format can improve the decoding accuracy under moderate-to-strong turbulence conditions. This work will be beneficial for the design of UOWC systems.