This thesis examines a number of underwater acoustic signals and the program of classifying these signals using a back-propagation neural network. The neural network classifies the signals it upon features extracted from the original signals. The effect on classification by using an adaptive line enhancer for noise reduction is explored. To feature extraction methods have been implemented; modeling by an autoregressive technique using the reduced-rank covariance method, and the discrete wavelets transformation. Both orthonormal and non-orthonormal transforms are considered in this study. Results show that best performances are obtained when using the non-orthonormal transform with multiple voices to allow for finer frequency partitioning of the energy features.

Back to homepage.