| dc.description.abstract | Marine environmental information is vital for marine engineering applications. Underwater acoustic remote sensing is a common approach to obtaining marine environmental information by acoustic waves. With the development of big data and high-performance computing, machine learning has shown superior performance in many research fields. This trend also attracts acoustic researchers to introduce machine learning to the community. Recently, most existing machine learning-based works in the field follow the supervised learning scheme whose superiority has been demonstrated given the vast labeled data. Unfortunately, the insufficiency of labeled data is a common problem in underwater acoustics since the approaches to get labels could be hard and/or high cost. For instance, ’coring’ is a direct but highly cost approach to get accurate local seabed acoustic parameters. The aforementioned limitation hinders the application of machine learning in real-world underwater acoustic remote sensing tasks. This thesis focuses on two underwater acoustic remote sensing tasks, i.e., source localization and geoacoustic inversion.
The first task is source localization which is to estimate the source location based on the received acoustic signal. The classic way for source localization is matched field processing, which only performs well without mismatch conditions. Due to the less sensitivity to mismatch conditions, researchers have introduced machine learning, especially purely supervised learning, for source localization. Most of the existing works concentrate on utilizing different state-of-the-art architectures of deep neural networks trained on a vast simulation dataset with the supervised learning scheme. However, this approach is not suitable to process a system collecting live acoustic signals (without labels) since the model trained with the purely supervised learning scheme needs labels. Focused on the capability of processing live acoustic signals, a self-supervised source localization framework is proposed and it is suitable for a real-world ocean monitoring system. The framework consists of a feature extractor based on contrastive predictive coding and a source localizer based on a multilayer perceptron. The localization framework is assessed by a public dataset and a proprietary dataset. The public dataset is obtained from the SWellEx-96 Experiment conducted in San Diego in 1996. The proprietary dataset is collected from a sea trial conducted in Trondheim Fjord in September 2022. The results on the public dataset show that the localization framework is superior to alternative methods in terms of localization accuracy, the robustness of receiver-depth selection, and generalization capabilities on unseen data. The proprietary dataset is collected by an acoustic vector sensor system which provides azimuth information. An azimuth estimation method is further adopted to calculate the azigram. The results on the proprietary dataset demonstrate not only the performance of the proposed localization framework but also the benefit of the acoustic vector sensor for source localization and azimuth estimation.
The second task is geoacoustic inversion which is to estimate the geoacoustic parameters of the marine acoustic waveguide based on the received acoustic signal. The prevalent approach for geoacoustic inversion is also matched field processing, which exploits an optimization method to iteratively search in the parameter space and find a set of geoacoustic parameters best-fitting the observed data. However, conventional optimization methods have a fixed search strategy which makes the method not effective enough in some cases. To make the inversion process faster and more efficient, an intelligent geoacoustic inversion framework based on reinforcement learning is proposed. More specifically, a deep-Q network is adopted as the whole framework for learning search strategy and controlling the agent to search effectively in the parameter space. Additionally, the environment and agent configurations of the framework are carefully designed for geoacoustic inversion. The inversion framework is tested on several sets of dispersion curves for estimating shear wave velocity profiles in the seabed. It is demonstrated that the proposed inversion framework performs a faster and more accurate inversion. | en_US |
