Feature based neural network acoustic transient signal classification.

Utilization of neural network techniques to recognize and classify acoustic signals has long been pursued and shows great promise as a robust application of neural network technology. Traditional techniques have proven effective but in some cases are quite computationally intensive, as the sampling rates necessary to capture the transient result in large input vectors and thus large neural networks. This thesis presents an alternative transient classification scheme which considerably reduces neural network size and thus computation time. Parameterization of the acoustic transient to a set of distinct characteristics (e.g. frequency, power spectral density) which capture the structure of the input signal is the key to this new approach. Testing methods and results are presented on networks for which computation time is a fraction of that necessary with traditional methods, yet classification reliability is maintained. Neural network acoustic classification systems utilizing the above techniques are compared to classic time domain classification networks. Last, a case studs is presented which looks at these techniques applied to the acoustic intercept problem.
http://archive.org/details/featurebasedneur1094524169
Lieutenant Commander, United States Navy
Approved for public release; distribution is unlimited.