Modulation Classification using Joint Time and Frequency-domain Data

In this paper we propose a modulation classification scheme based on deep learning and convolutional neural networks. The proposed solution is based on I/Q amplitude samples that are classified into a set of 24 modulations, including analog and digital modulations. We investigate the adoption of different features in the training data. The features are computed using I/Q amplitude samples and include statistical information obtained in the time-domain and frequency-domain. Given the high amount of features considered in the training data, we investigate the possibility of projecting the features' data into a set of uncorrelated variables through the Principal Component Analysis procedure. We evaluate the classification performance for the different modulations and signal-to-noise ratio (SNR) values. Moreover, we also quantity the performance of the classifier when the features' data is projected into a subset of the uncorrelated variables. Finally, we compare the performance of the classifier with other works already available in the literature, assessing the effective performance gains of the proposed solution.