Improving Data-Driven RF Signal Separation with SOI-Matched Autoencoders

While the use of deep learning-based methods in radio frequency (RF) signal processing has steadily increased in recent years, little attention was paid to RF signal separation, especially for scenarios with a single antenna receiver. In order to further investigate single-channel signal separation, the ICASSP 2024 SP Grand Challenge on "Data-Driven RF Signal Separation" was organized. This paper presents the challenge submission that was labeled LHen. We extend the WaveNet baseline model with an autoencoder that is matched to the signal of interest and significantly improves system performance in terms of mean squared error evaluation metric.