Radio Modulation Classification Using Deep Residual Neural Networks

We propose a new deep residual network for Automatic Modulation Classification, OPResNet-18. It achieves state-of-the-art accuracy on the RadioML 2016.10a data set. We train the proposed model and other state-of-the-art networks with augmented data by adding a Carrier Frequency Offset (CFO). We find that the previously proposed IQNet-3 is robust to CFO. We demonstrate that this robustness allows the performance of IQNet-3 to be further improved through data augmentation in contrast to existing neural networks that cannot handle CFO. Finally, we provide evidence that standard data pre-processing techniques for time-domain data that reportedly perform well in many domains do not perform as well as a simple alternative, the outer product, in the IQ domain.