Path optimization for acoustical oceanography applications

Abstract

This Master’s thesis contributes to underwater acoustic sampling techniques through machine learning. It has two objectives: (1) Optimizing the sampling process for underwater sound fields and(2) Optimizing data assimilation for ocean acoustic tomography. To address the first objective, we developed an approach that leverages autonomous underwater vehicles (AUVs) to sample unknown sound fields. Unlike fixed sensor networks with spatialconstraints, AUVs can make real-time decisions and adaptively survey a region. The proposed algorithm, sound exploration with active learning (SEAL), adaptively samples a survey region based on the sound field characteristics. SEAL uses an active learning strategy based on Gaussian Process (GP) regression to characterize a static sound field in a survey region. With each location sampled, the algorithm employs a GP to estimate the field and quantify the uncertainty in the predicted sound field. The uncertainty metric is used to choose the next sampling location. This dynamic approach maximizes the information gained by the AUV at the locations that it samples. SEAL also ensures efficient convergence toward the true distribution of underwater static sources in the sample region. Our algorithms were developed via simulation and were validated with a controlled experiment in a swimming pool [work funded by the NSF AI Institute in Dynamic systems and Catalyst t Awards for Science Advancement]. For the second objective, we aimed to optimize the process of integrating acoustic data intoocean models via ocean acoustic tomography. Ocean acoustic tomography (OAT) derives water column properties from acoustic observations. OAT traditionally uses ray tracing and requires making frozen ray approximations that can be limiting in some cases. Another approach involves iteratively updating sound speed profiles and re-running sound propagation models until the modeled travel times agree with measured travel times. However, this approach is computationally expensive. To optimize the iterative approach, we developed a machine-learning pipeline to map perturbations in sound speed profiles to corresponding changes in acoustic ray paths. The 2010–11 North Pacific Acoustic Laboratory (NPAL) Philippine Sea experiment dataset was used to develop and validate a neural network. To constrain the model’s learning to small perturbations in ocean states and observed acoustic travel times, sound speed profiles were decomposed using empirical orthogonal functions, with principal components used for training, while ray paths were represented as Fourier functions. The proposed neural network focuses on the variability in sound speed profiles and ray paths so that a predicted decomposed ray can be obtained for small changes in the ocean state [work funded by the Office of Naval Research].