YIP -Data Fusion for Broadband Geoacoustic Inversion
Abstract
Abstract: The proposed research aims to characterize the undersea environment based on a large number of heterogeneous acoustic information sources. The considered scenario consists of various sensors that record acoustic data in a shallow water environment. Broadband acoustic signals, including noise emitted by large ships, breaking waves, or marine mammals, will be used for geoacoustic inversion (GI).The methodologies developed in this project will consistently combine or #fuse# acoustic data from different information sources in a Bayes framework and at a very attractive computational cost. Information sources can either be raw acoustic sensor data, preprocessing results of a linear filter, or the result of a GI preprocessing stage. Bayesian estimation will embed acoustic models of undersea sound propagation. While a large variety of methodological approaches for GI are available in the literature, little attention has been paid to data fusion and scalability of computational complexity. This project will push the envelope of undersea environment characterization by addressing these limitations. It relies on the framework of graph-based Bayesian estimation and will develop critical innovations in the following three research areas (RAs).# RA 1, #Fundamental Limits of Data Fusion for GI# will apply recent results on Fisher information analysis to the uncertain ocean acoustic environments of real GI experiments. The resulting uncertainty characterization will provide essential insights into error sensitivity and error propagation in GI problems.# RA 2, #Model-Based Data Fusion for GI# will introduce graph-based Bayesian methods for undersea environment characterization that can consistently combine information provided by heterogenous data sources and that are scalable in the sense that their performance improves with a growing number of sensors while their computational cost remains feasible.# RA 3, #Learning-Aided Data Fusion for GI# will introduce truly novel GI methods where operations performed by graph-based Bayesian estimation are enhanced by a graph neural network (GNN). The GNN is either trained with annotated real data or synthesized data from undersea acoustic models.The main innovation of this project is that GI based on heterogeneous data sources will be performed in an overarching Bayesian framework combined with deep learning. This holistic approach is expected to improve undersea environmental characterization. Future Naval Relevance: Innovation resulting from this project are novel methods for GI that substantially improve the performance of marine sensing systems and thus lead to tangible advantages across multiple naval applications. In particular, the proposed research aims at developing capabilities that can potentially be implemented at large scales and provide an undersea environmental characterization at high resolution.Datasets: For performance evaluation, we will use datasets from ONR-funded projects. In particular, wewill use data collected during the shallow water evaluation cell experiment 1996 (SWellEx-96), the shallow water experiment 2006 (SW-06), and the seabed characterization experiment 2017 (SBCEX-17). These datasets are made available to the PI by the Marine Physical Laboratory. In all experiments, different acoustic arrays were deployed. In addition to broadband acoustic noise, e.g., as emitted by large cargo ships, tonal and broadband acoustic signals are projected by acoustic sources that have been towed by a research vessel.Approved for Public Release
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Mar 03, 2023
- Source ID
- N000142312284
Entities
People
- Florian Meyer
Organizations
- Office of Naval Research
- United States Navy
- University of California, San Diego