Classification of Digital Modulation Schemes Using Multi-layered Perceptrons

Automatic classification of modulation schemes is of interest for both civilian and military applications. This report describes an experiment classifying six modulation schemes using a Multi-Layered Perceptron (MLP) neural network. Six key features were extracted from the signals and used as inputs to the MLP. The approach was similar to that of Azzouz and Nandi [2]. The aim was to see how the performance of the classifier varied according to different neural network sizes and signal-to-noise ratios (SNR) ranging from 5 dB to 25 dB. It was shown that the performance degraded significantly for SNR of 5 dB for all network sizes. This was suggested to be due to a high noise sensitivity on some of the features. The best classifier had a success rate of 63.7% for 5 dB SNR and 94.8% and over for signals with SNR ranging from 10 dB to 25 dB. A property of neural networks are that they not necessarily have an all-or-nothing output. It was found that for a high proportion of wrongly classified signals the second strongest MLP outputs represented the correct modulation scheme. This information could thus be used to do an informed guess about alternative modulation schemes if the original classification is found to be incorrect.