| Passive sonar systems capture sound radiated in a one-way transmission from the transmission source to the receiver. The automated classification of passive sonar signals is a particularly challenging time series recognition task due to the complexities introduced by the ocean, their transmission medium.;Hidden Markov Models (HMM) have been widely used in time series analyses in general and speech recognition applications in particular. In this thesis an HMM classification system is developed to discriminate among fifteen classes of passive sonar patterns, comprised of both natural and man-made underwater signals, that were provided by the Defence Research Establishment Atlantic (DREA). The HMM-based classifier correctly labelled 91.3% of training patterns and 91.5% of test patterns. The performance of the HMM set is compared to the results of a finite impulse response neural network (FIRNN) that was trained for the same classification task.;In continuous-mode operation, the classifier must decide whether or not a pattern is present in addition to deciding to which class a detected pattern belongs. A methodology for using a set of HMMs to both detect and classify passive sonar signals is developed. As well, terminology and performance measures are introduced to describe and evaluate the continuous-mode operation of a sonar detection and classification system. |
