부산대학교 도서관

Trained submarine operators identify threats through passive sonar systems by analysing the acoustic waves captured by arrays of hydrophones. Automatic Classification Systems may be quite beneficial to address this task; however, typical operational settings require instruments to identify the occurrences of unknown classes of ships for promptly alerting this specialised crew. This article proposes a new approach for developing an accurate novelty detector of unknown classes of ships. This proposal comprises an architecture composed by a synergistic combination of cluster-specialised and bagging generated novelty detectors, in opposition to a previous solution exploiting a hierarchical class-specialised architecture. The proposed approach is experimentally evaluated using radiated noise from 8 classes of ships, acquired in an acoustic range from the Brazilian Navy. The proposed detector outperforms previous works in 4.4% of AUC (on average), assuming an evaluation scenario composed by five known and three supposedly unknown ship classes. This gain is relevant to this performance-critical application and is related to a more strict definition of class boundaries attained by the proposed strategical clustering and bagging combination.