As federal agencies collaborate with industry to refine standards and requirements for detection of federal incumbent radar signals in the 3550-3650 MHz portion of the Citizens Broad Radio Service (CBRS) band, there is a need to understand potential performance limitations of Evironmental Sensing Capability (ESC) detectors and to develop sound methods for performance evaluation.
Following up on the recently completed NASCTN “Waveform Measurements of Radars Operating in the 3.5 GHz Band” project, NASCTN has begun applying deep learning algorithms to collected spectrograms to better characterize factors that complicate detection of federal radars in the 3.5 GHz band. Using advanced methods of classification such as convolutional neural networks and long short-term memory recurrent neural networks, we have begun to explore the possibilities of improving upon methods of assessing channel occupancy, such as energy detection. By being able to accurately detect channel occupancy on large amounts of 3.5 GHz band spectrogram data, we can create a more accurate picture of channel occupancy and the expected power of out of band emissions when the band is not occupied. This information will be useful for determining optimal technical solutions and policies for sharing the 3.5 GHz band.
For this and future NASCTN projects involving machine learning, NASCTN has acquired high-performance workstation (pictured) specifically designed for deep learning applications.