Passive Sonar Classification Using Time-Domain Information and Recurrent Neural Networks

Abstract

Sonar systems have widely been used in both military and civilian applications. In particular, passive sonar systems play an important role in submarine operations in any nation’s Navy. Usually, passive sonar signal processing is performed in frequency domain for target detection and identification. Alternatively, in this work, a classifier based on recurrent neural networks and fed from the time-domain information is proposed. The proposed model employs Long Short-Term Memory (LSTM) networks aiming at classifying signals coming from 24 classes of military ships, which were organized into 4 super-classes based on expert knowledge. The model achieved an accuracy of 86.03%±3.08% outperforming a multilayer perceptron network (MLP) baseline model that was fed from frequency-domain data and obtained from Short-Time Fourier transformation.