Novelty detection in passive sonar systems using a kernel approach

In naval warfare operations, several techniques have been developed for passive sonar signal detection and classification. Sonar systems operate over very noisy conditions and, in modern warfare scenario, it might be necessary to classify ships that were not available for the classifier training process. Kernel-based algorithms efficiently access high-order statistics and, because of this, they can be used as preprocessing and classification techniques. Support vector machines (SVMs) are one of most common supervised kernel-based learning models and one of its applications is one-class SVM, which detects events that were generated from the same generating function estimated along the training process. Kernel PCA (kPCA) is kernel-based extension of principal component analysis (PCA). This paper proposes the application of experimental sonar data to one-class SVM model combined with kPCA to detect ships events that were not available in the training process, i.e. novelty class events.