Deep Architectures for Modulation Recognition with Multiple Receive Antennas

Modulation recognition using deep neural networks has shown promising advantage over conventional algorithms. However, most existing research focuses on single receive antenna. In this paper, modulation recognition with multiple receive antennas using deep neural networks is investigated and four different architectures are introduced, including equal-gain CNN, multi-view CNN, 2-dimensional CNN and 3-dimensional CNN. Each architecture is constructed based on a ResNet and tuned to the extent that its performance does not further improve when the network size and parameters change with a given dataset. These architectures are then compared in terms of classification accuracy. Simulation results show that 3-dimensional CNN yields the overall best performance, while the equal-gain CNN leads to the lowest performance. Further, both 2-dimensional CNN and 3-dimensional CNN, which jointly extract features from multiple receive antennas with different feature encoding, outperforms either equal-gain or multi-view CNN, which fuses extracted features from each antenna. This indicates that utilizing inherent structures within deep neural networks to jointly extract features from different antennas can achieve better performance than the schemes that combine individually encoded features from each antenna, and extending the dimension of CNN from two to three can enhance feature extraction capabilities in the context of modulation recognition.