Autoencoders for Signal Enhancement in Communication Systems

Autoencoders are particularly interesting deep learning models for communications, as they resemble the architecture of a classical transmission system. Several works explored the idea to learn the transmitter, i.e., an encoder neural network and the corresponding receiver (decoder neural network) jointly in an end-to-end manner. We propose a novel hybrid autoencoder system that amends a conventional transmission chain. The autoencoder inputs the modulated signal of a conventional transmitter and adds desired features to it, e.g., a possibly constant signal envelope and an improved resistance against transmission errors. This is achieved by the choice of proper loss functions during training and by adding suitable regularization layers to the encoder. The decoder part at the receiver aims to reobtain the classically modulated signal, such that the proposed signal enhancing autoencoder is transparent to the conventional transmission system. Novel features can be added to an existing transmitter-receiver pair. Our presented results show that the proposed autoencoder can improve existing communication schemes in versatile ways with very low efforts. The proposed concept is particularly useful in the military context, where communication systems are deployed in large numbers and have long utilization periods, but might need improvements or novel features over their lifetime.