Adapting Deep Learning for Underwater Acoustic Communication Channel Modeling

Abstract

Listen

Translate article

The recent emerging applications of novel underwater systems lead to increasing demand for underwater acoustic (UWA) communication and networking techniques. However, due to the challenging UWA channel characteristics, conventional wireless techniques are rarely applicable to UWA communication and networking. The cognitive and software-defined communication and networking are considered promising architecture of a novel UWA system design. As an essential component of a cognitive communication system, the modeling and prediction of the UWA channel impulse response (CIR) with deep generative models are studied in this work. Firstly, an underwater acoustic communication and networking testbed is developed for conducting various simulations and field experiments. The proposed test-bed also demonstrated the capabilities of developing and testing SDN protocols for a UWA network in both simulation and field experiments. Secondly, due to the lack of appropriate UWA CIR data sets for deep learning, a series of field UWA channel experiments have been conducted across a shallow freshwater river. Abundant UWA CIR data under various weather conditions have been collected and studied. The environmental factors that significantly affect the UWA channel state, including the solar radiation rate, the air temperature, the ice cover, the precipitation rate, etc., are analyzed in the case studies. The obtained UWA CIR data set with significant correlations to weather conditions can benefit future deep-learning research on UWA channels. Thirdly, a Wasserstein conditional generative adversarial network (WCGAN) is proposed to model the observed UWA CIR distribution. A power-weighted Jensen–Shannon divergence (JSD) is proposed to measure the similarity between the generated distribution and the experimental observations. The CIR samples generated by the WCGAN model show a lower power-weighted JSD than conventional estimated stochastic distributions. Finally, a modified conditional generative adversarial network (CGAN) model is proposed for predicting the UWA CIR distribution in the 15-minute range near future. This prediction model takes a sequence of historical and forecast weather information with a recent CIR observation as the conditional input. The generated CIR sample predictions also show a lower power-weighted JSD than conventional estimated stochastic distributions.