Deep Learning-Based Demodulation in Impulse Noise Channels

Wireless systems on military platforms often operate in challenging electromagnetic environments. Such environments contain interference which tends to be non-Gaussian. A more suitable model to derive the necessary receiver algorithms is the Symmetric a-Stable distribution. Adapting a receiver to the a-Stable model provides significant performance gain in challenging interference environments compared to Gaussian-based receivers, at the cost of higher computational complexity. With the emergence of deep learning and neural networks, a promising application is replacing the steps of demodulation with a suitable network structure in order to decrease the complexity. In this work, the generalization capability of such networks is explored in order to examine the models' ability to adapt to new, unseen error correcting codes of different length and type, and if good performance can be achieved in comparison to traditional demodulation as well as if the networks will consider code structure when demodulating. The results show that a networks ability to generalize largely depends on architecture, as well as training data. Furthermore, the models which were unable to generalize made bad decisions based on code structure assumptions, resulting in worse performance compared to the more general models even though it was trained and tested on the same code. It is shown that the proposed model performs well, on par with the a-Stable based method, with significantly lower computational costs.