Localização de fontes submarinas por processamento de campo casado, focalização e aprendizagem de máquina

Abstract

Source localization is a classic inverse problem of great interest in underwater acoustics. Matched field processing (MFP) combines the processing of a received signal with replicas of the acoustic field calculated by a propagation model to invert for source range and depth, thereby exploiting the complexity of the full acoustic field in a waveguide. Although a versatile technique with multiple applications in acoustic inversion, MFP is limited by the environmental model used to describe the waveguide. If the model is not an accurate depiction of the waveguide, a problem called model mismatch arises, with implications to the technique’s robustness. This dissertation explores the use of two methods developed as more robust alternatives to classic MFP. The first one, focalization, is an adaptive extension of MFP that treats inversion as a global optimization problem by including other model parameters in the search space. A search heuristic, in this work based on genetic algorithms, is used to find the global maximum of an objective function. The heuristic is used to locate in depth and range the source of an experimental dataset in an uncertain environment with low a priori knowledge of its parameters. Results show significant improvement over classic MFP, taking the accuracy from 31% to 73%. Another approach is developed based on machine learning with a focus on the use of modeled fields for neural network training. It is demonstrated that a neural network is capable of approximating the processor used in MFP and that, when environment uncertainty is considered in the training of the neural network, it becomes more robust and produces more consistent source range estimations than MFP, obtaining only 2.96% error for the dataset used.