Constrained neural networks for recognition of passive sonar signals using shape

Abstract

The author describes a neural network system that recognizes seven different types of passive sonar signals from their characteristic shapes. The system has a preprocessor for signal detection and symbolic representation, a bank of three highly constrained feedforward neural networks for recognition, and a postprocessor for network interpretation and performance adjustment. The preprocessor uses image processing and morphological techniques to extract and track energy, and converts each detected signal into a chain code. The chain code is passed to an ensemble of three independent neural networks, each of which votes on the signal's type. The system's performance on 1400 unseen test signals was an adjustable 93% overall correct recognition rate, 5% error rate, and 2% rejection rate.<>