Classificação neural hierárquica de sinais de sonar passivo com exploração da fusão das informações temporal e espectral

Abstract

The acoustic noise radiated by ships propagating in the seas is detected and classified by military submarines through passive sonar system techniques. This system is also used for monitoring the Brazilian coast and in applications in the civil environment. Due to the diversity of existing ships, interference from ambient noise and the complexity of identification of contacts, thus the development of systems that perform automatic identification and classification using processing algorithms of signals and computational intelligence have been gaining ground in the area of military and academic research. Many works in the literature address the development of ship classifiers trained with spectral information and with few targets involved, which is far from the reality in which we face a great number of classes and signals in the time domain. This amount of classes also becomes a challenge for a classifier. Thus this work proposes the hierarchical classification of passive sonar signals, exploring the fusion of temporal and spectral information through two analyzes (LOFAR and Temporal) and the use of deep neural networks (more specifically convolutional and recurrent) in order to make the classification of sonar system better. In this analysis two experimental datasets were used: the first acquired from submarine in operation and the second from a passive sonar monitoring the maritime traffic of a port region of Rio de Janeiro. The results show that the proposed method and the recurrent network present good results, 88, 76% ± 2, 81% and 85, 55% ± 3, 28% respectively in comparision with traditional neural networks.