Basis Expansion with Transformer and Two-map Spectrum Inference (BETTSI)

Shared Spectrum Company (SSC) and George Mason University (GMU) propose to design Basis Expansion with Transformer and Two-map Spectrum Inference (BETTSI). BETTSI is a distributed, coherent sensing solution to generate a spectrum map of available channels in sparse or dense spectral environments for channel allocation in a decentralized multi-hop network. Sensing-based spectrum allocation is a difficult problem due to the large problem space and the lack of closed-form solutions.  This is compounded by the probabilistic nature (propagation effects, unknown antenna patterns, etc.) of RF interference. The large problem space and the ability to model interference suggests a Machine Learning (ML) solution. The BETTSI approach has multiple technical innovations.  The ‘Transformer’ approach is one of the innovative elements in the proposed GMU distributed ML spectrum inference approach. The concept of “transformer” is a relatively new topic in AI/ML and is becoming more popular in various applications including Natural Language Processing.  GMU has extensive AI/ML experience applied to spectrum use problems. GMU has more than 20 published papers in the distributed sensing ML area. The problem is amenable to ML if accurate training data is available.  SSC has over 20 years of experience in modeling DoD/commercial system interference. Another challenge is the spectrum sensing ‘Hidden Node’ problem. In general, the sensor coverage is limited, and primary users located outside of the coverage area must be protected from interference. Forming Primary User maps from sensor data is insufficient. Secondary users located within the sensor coverage area can still cause interference to these primary users. Unless the Hidden Node problem is accounted for in the design, the system will cause significant interference to primary users and the system will not obtain approval from spectrum managers.  SSC has developed and field tested a Dempster Shafer ‘BDI Map’ approach (U.S. Patent 8,064,840) that determines the sensor coverage edge (‘ignorance’) that solves the Hidden Node problem.  The BETTSI sensors determine the signal’s frequency, amplitude and bandwidth. The sensor classifier potentially determines the signal angle of arrival and emission type. The sensor information (antenna gain, height, etc.) are known by the BETTSI system. The sensors also determine ‘null’ detections where no signals are found above a known power level threshold value. Null detections are critical to spectrum allocation decisions because the system goal is to estimate the size (spatial and frequency) of spectrum holes compared to the secondary user’s interference area. If the sensor doesn’t inherently make null reports, then the BETTSI system infers the reports based on a lack of detection reports and known sensor information (antenna parameters, sensitivity and frequency revisit rate).