Neural network clusters and cellular automata for the detection and classification of overlapping transient signals on radio astronomy spectrograms from spacecraft

Abstract

We address the problem of automatic detection and classification on spectrograms of mixed planetary low frequency radio signals with additive plasma noise. The signals and the noise under study are overlapping, nonGaussian, non stationary and non linear. The data obtained from spacecraft telemetry are irregularly sampled. We show a series of preprocessings that enables the use of neural networks. A cluster of time delay neural networks is then used to observe the signals from many windows. The different outputs of the time delay neural networks are the inputs of multi layer perceptrons which yield an intermediate classification. Cellular automata with a look up table of rules derived from the physical laws governing the radio electric phenomena do the find pattern recognition in a deterministic number of iterations.