Adaptive FSK decoding with an artificial neural network

Abstract

We describe an empirical study of the capability of an artificial neural network (ANN) to decode a frequency shift key (FSK) signal. An algorithm for generating a minimal, yet comprehensive ANN training data set is discussed. The FSK signal is over sampled. The samples are presented to the ANN as a window in time. The window is one symbol wide. After initial training, white Gaussian noise is added to the samples and the ANN's ability to generalize is tested. We then conduct additional training, using the noisy data, to test the ANN's ability to adaptively recover. Simulation results are reported.<>