| dc.description.abstract | This study investigated source localization in the ocean using machine learning meth-ods. By interpreting the sound waves emitted by the source, localization can be acquired.In order to interpret the pressure, the machine learning models support vector machines(SVM) and feed forward neural networks (FNN) have been considered. The performanceof the models have been compared, both when using regression and classification. Somespecific cases have been studied; How does a noisy pressure field affect the localization?How does a reduction in frequencies, hydrophones and pressure data size affect the local-ization?This work used a simulation framework to produce the pressure data used by the ma-chine learning models. The initial phase of this project consisted of processing the data, inorder to make it more comprehensible and adding different levels of noise. The machinelearning models, both FNN and SVM, has a set of free parameters, and much time wasdevoted to tune the models in order to acquire reasonable results.When the models were tuned, a lot of simulations were run, and some interestingresults were acquired. The FNN/SVM classifiers showed roughly the same performance,and outperformed the FNN/SVM regression models. A decrease in the Signal-to-noiseratio led to overfitting and worse performance. When decreasing the amount of frequenciesand the size of the data set, the localization was less accurate. Regarding hydrophones, lesshydrophones generally reduced the performance, but when using the hydrophones closestto the surface and the source, the performance was significantly better than the contrary. | en |
