Acoustic Classification with Neural Networks

Abstract : A major objective of the research that is described in this report was the preliminary development and testing of neural networks to aid a sonar operator in detecting and classifying sonar signals and signatures. Two networks were developed to recognize a set of sounds that is representative of sonar pulse type signals. The pattern vector inputs were derived from the signal spectra. Both networks are time delay neural networks that classify by recognizing spectral shapes and the first and second differences of spectral shapes. One is a back propagation network and the other is a network that does not contain weights and thus is much simpler than back propagation networks to implement in hardware. Both networks demonstrated a very low false alarm rate and were as accurate or significantly more accurate than the test subjects at signal detection and classification. When the subjects were prompted by a network they were more accurate at finding and classifying the more difficult to detect signals. Our preliminary conclusion is that neural networks will have great value in detecting sonar pulses which is an important signal class. The architecture and understanding of the weightless network that was developed during Phase I leads us to believe that networks of this type may become important in sonar signal analysis. (jhd)