Detecting coherent sources with deep learning

Detecting correlated sources in a dynamic radio frequency (RF) environment is both challenging and critical to antenna array processing. We introduce a deep learning framework capable of detecting both correlated and uncorrelated radio frequency sources in the presence of ambient noise and multiple interference signals. The auto-correlation matrix is extracted from the received signal matrix and spatially smoothed using forward-backward averaging. The processed signal is then used as an input to a ResNet34 architecture which detects the number of sources present in the sampled waveform. We transform the source detection problem into a one-vs-all binary classification problem where, the machine predicts a binary label corresponding to the number of detected sources. The designed framework is trained and evaluated on simulation data closely replicating real-time RF environments.