Preprocessing passive sonar signals for neural classification

Abstract

The noise radiated from ships in the ocean contains information about their machinery and can be used for detection and identification purposes. Here, a preprocessing method is developed in order to improve the performance of a feedforward neural network, which is used to classify four classes of ships. The entire system operates in the frequency domain over the information collected by the sensors of a passive sonar system. The effect of spectra averaging, resolution and background noise normalisation in the classifier performance is evaluated. Using preprocessed data to feed the input nodes of the classifier, a classification efficiency of about 97% has been achieved.