Distributed Machine Learning and Sensor Fusion for Spectrum Sensing System Optimization
Abstract
The growth in volume of programmable wireless devices, such as software defined radios (SDRs), is beginning to cause significant challenges in interference mitigation for both commercial and government radio frequency (RF) operators. Simultaneously, the advent of new wireless technologies and devices is creating spectrum scarcity. To enable seamless connectivity, developing solutions to share the spectrum among active and passive devices is critical. Current studies have mostly focused on developing spectrum sharing techniques across primary and secondary users in a given environment, assuming that perfect spectrum information of the environment is available. While these efforts are important building-blocks, there are fundamental gaps to translate such methods into practical systems that the government and military can rely on. In particular, developing the technologies for collecting compressed and quantized spectrum information from distributed sensors is essential. Such technologies will enable ubiquitous spectrum sensing that can be realized via low-cost RF sensors. Moreover, intelligent functions like signal classification, anomaly detection, and spectrum usage prediction can greatly optimize these environments. Machine learning based on the sensor data has the potential of providing this intelligence, but studies of machine learning in the spectrum domain barely exist today. In this program, we consider a network of low complexity RF sensor nodes that are able to sense the spectrum and collect samples from their environment. We propose developing novel link-level methodologies for quantizing, fusing, and compressing these measurements into efficient spectrum usage vector representations at distributed fusion nodes. We then propose novel methodologies for distributing spectrum machine learning tasks among the network of fusion nodes, including specific models for spectrum usage prediction and anomaly detection based on the vectorized RF samples. The outputs of these machine learning tasks are subsequently fed back to the commercial or government/military users for command and control purposes. Compared with the current state-of-the-art, our proposed program is expected to advance the spectrum fundamental research in two complementary directions: (i) compressed and quantized spectrum information gathering and fusion via low-complexity sensors, and (ii) distributed understanding of the spectrum environment using machine learning solutions at the fusion nodes. This architecture would avoid incurring the large communication costs and latencies associated with transferring the volumes of raw RF measurements collected by the sensors upstream to datacenters for processing. To enable this program and minimize the associated risks, we envision that our efforts will be guided by the database of signals emerging from the DARPA RF Machine Learning Systems (RFMLS) program that Crane played a critical role in developing.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- May 19, 2020
- Source ID
- N001642011004
Entities
People
- Christopher Brinton
Organizations
- Naval Surface Warfare Center
- United States Navy
- University of Virginia