Agile Communication Architectures
Despite years of intense research and much progress in spectrum sensing and signal classification, the vision of fully cognitive radio networks has not yet been realized. One of the main roadblocks towards the development of intelligent radios has been the complexity of learning the radio environment. The wide range of modulation formats and the diversity of channel impairments (noise, interference, fading, intersymbol interference, hardware imperfections) render the task challenging, especially if it has to be performed in real time. Recent advances in machine learning, especially the development of deep learning and Convolutional Neural Networks (CNNs), are enabling fast and accurate classification that promise significant improvements over traditional techniques.
It has been recently shown that deep learning offers a competitive alternative to state-of-the art systems for modulation recognition in wireless networks. The reported results show that a small number of time samples is already enough to accurately distinguish many modulation types. Motivated by these promising results, and based on our experience with CNNs in other settings, we investigate here a hybrid learning approach that also exploits known useful features to assist the CNN as well as using flexible time-space decompositions that are more in line with the actual learning task.
Figure 1. Network architecture combining conv nets with side-information, such as e.g., cyclostationary features
Figure 2. Classification accuracy on RadioML test set. From SNR > 0.0 and beyond classification accuracy is > 0.9.
Figure 3. Confusion matrix at SNR = 18 for RadioML test set.
Figure 3. Software-defined radio used for prototyping and benchmarking of collaboration approach.
Figure 4. As part of the challenge information has to be transmitted between two radios while sharing the spectrum with other participants