Application of neural-network for real-time underwater signal classification

A modified neural network architecture is introduced for the problem of classification of underwater acoustic signals in real time. The tasks involve signal analysis, feature extraction and classification of underwater acoustic signals. A great deal of effort has been made in frequency domain feature extraction and real time neural network classification algorithms. In order to detect the task-relevant part of the underwater signal from background noise and other interference, a fast and accurate spectrum estimation procedure was designed, which required a close study of signal pre-processing, segmentation, windowing, frequency estimation and post-processing. Then by use of the self-learning and adaptive processes of a neural network, the parameters of the network model can be justified. These parameters were utilized to organize the network model, which produced very good classification results. In order to meet the real time processing requirement, a neural network with a two step organization is proposed.