Enlarging neural class detection capacity in passive sonar systems

A neural discriminating analysis is used for classifying passive sonar signals. Preprocessed information from the amplitude spectra of the noise radiated from ships is projected onto only a few principal discriminating components for feeding the input nodes of the neural classifier. Envisaging practical applications, in which new incoming classes not known by the time of the training phase have to be detected in the production phase, a method is provided using the identification of outliers to trigger the arriving of a new class. Using experimental data, it is shown that up to 85% of patterns from an untrained class can be identified, without significant decrease on the efficiency of the classifier for classes known beforehand.