Title: Instance-based novelty detection in passive sonar signals
Authors: Victor Hugo Da Silva Muniz; João Baptista De Oliveira e Souza Filho; Eduardo Sperle Honorato
Addresses: Signal, Multimedia and Telecommunications Laboratory, COPPE, Federal University of Rio de Janeiro, Rio de Janeiro, Brazil ' Signal, Multimedia and Telecommunications Laboratory, POLI/COPPE, Federal University of Rio de Janeiro, Rio de Janeiro, Brazil ' Signal, Multimedia and Telecommunications Laboratory, POLI, Federal University of Rio de Janeiro, Rio de Janeiro, Brazil
Abstract: In submarines, sonar operators have the main task of identifying potential threats, named as contacts in the military jargon. The principal tool exploited when dealing with such situations is the passive sonar system. Automatic contact classification models may relieve the huge sonar operator workload but require mechanisms capable of identifying any contact not considered during system development. This paper discusses the development of a hierarchical instance-based detector of unknown contact classes for passive sonar signals, focusing on practical strategies for its hyperparameter tuning and performance assessment. Experimental data exploited in system evaluation comprises the acoustic noise irradiated by 28 ships belonging to 8 classes. These ships were submitted to different operational conditions in several runs conducted in an acoustic range. The kNN algorithm has performed best, achieving a novelty detection rate of 78.0%, associated with an average known case identification rate of 95.0%, considering a five unknown class evaluation scenario.
Keywords: passive sonar; novelty detection; automatic classification systems; instance learning; kNN; k-nearest neighbours.
DOI: 10.1504/IJICA.2022.124238
International Journal of Innovative Computing and Applications, 2022 Vol.13 No.3, pp.161 - 171
Received: 14 Apr 2020
Accepted: 21 Apr 2020
Published online: 19 Jul 2022 *