Classification of acousto-optic correlation signatures of spread spectrum signals using artificial neural networks

An attempt was made to determine if artificial neural networks (ANNs) can be trained to classify the correlation signatures of direct-sequence and frequency-hopped spread-spectrum signals. Secondary goals were to determine if network classification performance can be modeled with a conditional probability matrix; if the symmetry of the matrices can be controlled; and if using a majority vote rule over independently trained networks improves classification performance. Correlation signatures of the spread-spectrum signals were obtained from US Army Harry Diamond Laboratories. The signatures were preprocessed and separated into various training and testing data sets. Thirty samples of network responses for several sets of training conditions were gathered using a neural network simulator. ANNs trained directly on correlation signature data yielded classification accuracies on test data at or near 80%. The probability matrices were stationary with regard to test sets, and the ability to shift the symmetry of the matrices was demonstrated. Improvement of classification accuracy via majority vote was possible if the nets were trained on different data sets. An average improvement of 1.8% was found to be statistically significant for α=0.05. A metric was developed to estimate the similarity of the solutions found by networks in a given training run