On data-driven underwater acoustic direct localization: Design considerations of a deep neural network-based solution

Abstract

High reverberation, depth-dependent speed of sound and intricate environmental structures are a few of the challenges that make the development of reliable underwater acoustic localization (UAL) methods hard. Since the nontrivial physics of underwater acoustics and the inherent random effects amount to a highly complex statistical model of the measured data, it is natural to consider deep neural networks (NNs) as a central tool in the development of UAL methods. Indeed, recent studies have shown that some properly designed NNs can lead to unprecedented localization capabilities and enhanced accuracy. However, out of the already myriad available possibilities, it is not immediately clear how to choose a proper NN architecture—input structure, layer types, training procedure, loss function(s), etc.—that will lead to successful operation. In this talk, we will present an analytically informed architecture that can learn (for now, in simulations) to localize in complex underwater environments. We will also provide analytical arguments that can explain the robustness of such architectures to certain types of propagation/environment model mismatches. The operation of the discussed architecture will be demonstrated via simulation results in a rich, reverberant channel. The results suggest that deep NNs could become a viable solution to the UAL problem.