Classificação de navios com uso de sinais de sonar passivo

Abstract

The classification of ships by their acoustic noise radiated at sea is one of the essential tasks carried out on board of a submarine. Several ship classifiers have been proposed in the literature, but none of them has an operative focus, mainly in relation to the data used, and the significant number of classes that occur in an operating environment. In addition to the academic contribution of developing a classifier closer to the reality of a submarine, this work also meets one of the demands of the development of the Brazilian nuclear submarine, an autochthonous classifier. Four classifiers are designed: multilayer perceptron (MLP), class expert committees (one based on MLP and other based on Support Vector Machines - SVM) and a hierarchical committee which is built up on expert knowledge on vessels. The input data dimensionality reduction is performed through the Principal Components of Discrimination. The classifiers were applied to two experimental databases: one from recordings of acoustic signals collected from passive sonars of submarines in operation (with 24 classes) and another from recordings of acoustic signals collected from a sonar passive platform monitoring the maritime traffic in a port region (with 31 classes).