NIST taps AI for better radar detection in 3.5 GHz band

Researchers at the National Institute of Standards and Technology (NIST) say they’ve come up with a better way to detect offshore radars that could help commercial users know when they need to yield the 3.5 GHz band to federal users—and it involves a form of artificial intelligence (AI).

In a new paper, NIST researchers demonstrate that these deep learning algorithms are significantly better than a commonly used, less sophisticated method for detecting when offshore radars are operating.  

NIST has been involved in the development of standard specifications to enable commercial users to operate in the 3.5 GHz band, aka the Citizens Broadband Radio Service (CBRS) band, while assuring the U.S. Navy that the band can be successfully shared without RF interference. The Wireless Innovation Forum Spectrum Sharing Committee, the public-private standards body for CBRS, approved the specifications in February 2018.

Currently, there are no official standards for determining when the military is actually using the band. Radar signals from ships at sea are detected using automated detectors that look for energy rises in the electromagnetic spectrum, but these detectors are not discriminating enough to consistently get it right, sometimes confusing other RF signals as radar or missing the radar signatures altogether, according to NIST researcher Michael Souryal, who is one of the authors of the new paper.

RELATED: FCC’s 3.5 GHz CBRS proposal sparks cheers, fears

So Souryal and his colleagues turned to AI for a potential solution. Eight deep learning algorithms—software systems that learn from pre-existing data—were trained to recognize offshore radar signals from a collection of nearly 15,000 60-second-long spectrograms. These spectrograms were recorded in 2016 near naval bases in San Diego, California, and Virginia Beach, Virginia, for NIST’s National Advanced Spectrum and Communications Test Network, according to NIST

After training, the deep learning algorithms were pitted against energy detectors to see which performed best at identifying and classifying a set of spectrograms different from the ones used to educate the AI detectors. The researchers found that three of the deep learning algorithms appreciably outperformed the energy detectors.

While the research sounds promising, it's not expected to have an immediate impact on how radar is detected in the CBRS band. It's not related to any proposed rule changes at the FCC but simply informs future activities.

“More specifically, results of this work could inform the community of current usage of the band (its occupancy and background interference levels, at least at the locations and for the time periods observed), and related ongoing research on AI detection of 3.5 GHz signals can inform commercial development of sensors," Souryal said in a statement provided to FierceWirelessTech.

The NIST researchers said they plan to continue refining the AI detectors by training them with higher-resolution, more-detailed radar data, which could lead to even better performance.