ASA 124th Meeting New Orleans 1992 October

4aUW15. The performance of artificial neural networks in classifying acoustic signatures.

Fred C. DeMetz

15455 Glenoaks #325, Sylmar, CA 91342

The performance of feedforward networks employing the backward propagating delta rule for error correction have been tested utilizing simulated acoustic target signatures and noise. The network correct classification and false alarm rates were studied for varying size and composition of the training sets. The effects of signal-to-noise ratio, and fluctuations in the amplitude and frequency of the signal data were also investigated. The practicality of utilizing networks in multistage sonar processors is assessed from the standpoint of normally available training set size, composition, and quality.