A Neural Network Model for Assisting Project SERENDIP
Denise M. Taylor, Michael Towsey, Joachim Diederich · 1998
Project SERENDIP currently uses a signal-to-noise ratio threshold method for the extraction of features from power spectra for further analysis. This paper details the results of a small-scale experiment conducted to test the viability of replacing this method with an artificial neural network approach. A simple recurrent network is trained to recognise simulated extra-terrestrial signals in a time-frequency sequence. The output of the network is used as a mask to extract the appropriate features from power spectra. Simulation results for the network integrated into a simple receiver model demonstrate the potential advantage of this approach over current practice.