Application of real data to sonar detection by neural networks

Abstract

This memorandum describes the results of applying an artificial neural network for the detection of target echoes. The present study is a continuation of work conducted in 1989 in which a three-layer artificial neural network was used to detect target echoes in simulated sonar data. The use of real data here has made it necessary to investigate the influence of various network-dependent parameters, such as learning rate and other parameters that describe noise conditions. In an attempt to speed up the learning process and to improve performance, a network architecture that contains direct connections from input to output is investigated. This modified architecture allows the number of hidden nodes to be reduced while still maintaining the same network

performance. The real data include 66 received signals, 16 of which contain clear target detections and are used for training. The remaining 50 form a test set. The detection performance of the network for the test set is characterized by a probability of detection Pdet = 0.84 and a probability of false alarm

Pfa = 0.14.