Machine Learning Based Underwater Optical-Acoustic Communications Channel Switching for Throughput Improvement

Underwater Wireless Optical Communication (UWOC) is at the cutting edge of subsea networking, offering high capacity, low-latency, energy-efficient connectivity that offers many advantages over the acoustic standard which has been embedded in submersible systems for a century. One aspect in which it fails currently, however, is transmission range and reliability, only achieving 10s of metres in range and requiring Line of Sight (LOS) to operate, meaning that changes of turbidity and ambient noise originating from the Sun or ROV light sources can actively interfere with transmission success. An investigation into machine learning algorithms has been carried out that aimed to enable a modem to utilise environmental sensors to interpret the UWOC channel and make accurate predictions on whether it should transmit, potentially store in memory for later transmission, at the cost of latency, when the channel is clearer, or use another mechanism such as acoustics or radio frequency to transmit promptly, with minimal latency. It was found using a synthesized dataset compiled using simulation and a regarded photon-counting model, that common ML algorithms such as Support Vector Machines (SVM), Random Forest (RF) and Narrow-Neural Networks (N-NN) can successfully use parameters such as distance, transmission power and extinction coefficient to determine the nature of the channel, thus, whether to transmit or not, with classification accuracies greater than 98.5% providing a reliable method to switch between acoustic and optical signalling in response to channel conditions in the latter, maximising data throughput, reliability whilst managing energy consumption and latency.