A Hierarchical Ensemble of LSTM-based Autoencoders for Novelty Detection in Passive Sonar Systems

Sonar operators represent a vital workforce for identifying potential threats to submarines (referred to as contacts) by analysing underwater acoustic signatures acquired by their passive sonar systems. Automatic contact classification models may alleviate the sonar operator task but require additional tools for identifying any class of contact not considered during the system development. This paper proposes a hierarchical detector of unknown contact classes for passive sonar based on modelling signal spectra using Long Short-Term Memory Autoencoders networks. Considering the radiated noise of 28 ships belonging to 8 classes acquired in the Brazilian Navy acoustic range, the system achieved an expressive average value for the area under the detection operation curve (0.946) in a simulated novelty detection scenario involving five known and three unknown classes, surpassing the state-of-the-art.