Knowledge Embedding Networks based on Deep Learning for Automatic Modulation Classification in Cognitive Radio

Deep learning has shown remarkable success in cognitive radio. However, popular approaches mainly focus on the purely data-driven architecture design, and fail to explore the professional knowledge of wireless communication which is particularly significant for radio signal essential feature extraction. Inspired by digital signal processing theories, we propose knowledge embedding networks (KENs) to introduce the high-order statistics and radio spectral information into deep neural networks, based on the high-order convolution and multi-spectral attention mechanism. KENs are the general form of attention mechanism from frequency perspective and leverage identical structures of the conventional convolution layer for high-order statistics. To extract efficient representations of radio signal, frequency loss with generative adversarial networks is proposed to enhance the discrimination and richness of attention knowledge in an explicit manner. KENs enhance interpretability concerning essential radio properties while achieving the state-of-the-art accuracy by a significant margin compared to traditional automatic modulation classification approaches on RADIOML 2018.01A dataset.