A neural network harmonic family classifier

Abstract

A method of automatic target discrimination and classification for passive sonar applications is obtained using an unsupervised, self-organizing, Kohonen neural network. Utilizing detected, smoothed spectral data, the kth-order frequency differences are histogrammed into constant percentage resolution spectral bins prior to presentation to the network. This nonlinear transformation results in good neural network classification when the input data is topologically close to the training set. Simulations of the algorithm show highly robust performance to missing spectral detections, spurious spectral detections and Doppler shifts.