An Effective Convolution Neural Network for Automatic Recognition of Analog and Digital Signal Modulations for Cognitive SDR Applications

This paper proposes a convolution neural network (CNN) architecture for automatic recognition of signals on the basis of their modulations for Cognitive software defined radio (SDR) applications. It is developed starting from two CNNs specifically designed for this problem and is characterized by having a low number of convolution and fully connected layers sharing also a very low number of filters/units. Moreover, Batch normalization is used to increase learning rate and reduce training time. The reduced complexity together with its low operating time make it compliant with real-time SDR applications. The proposed architecture is validated on the RadioML2016.10a dataset showing interesting results in discriminating both analog and digital modulations under different signal to noise ratio (SNR) regimes.