Novelty Detection in Passive Sonar Systems using Stacked AutoEncoders

Modern military ships have become more silent in the recent years, so that their detection, tracking, and classification have become even more challenging. Submarines rely very much on passive sonar systems, which might operate in a multi-class scenario. The training data for (semi) automatic decision may not contain all possible classes; i.e., classifiers will face unknown classes in operation. Here, Stacked AutoEncoders (SAE) are applied for such novelty detection. Beyond that, the impact on classification efficiency is estimated, and a comparison with a shallow learning algorithm is performed.