PHYSICS-AWARE MACHINE LEARNING FOR DATA-DRIVEN, COGNITIVE RF SENSING
Abstract
Despite advances in Artificial Intelligence and Machine Learning (AI/ML) and signal processing for automatic target recognition (ATR) using surveillance radar, there remain significant challenges to robust and accurate perception in operational environments, such as that encountered for base security and perimeter defense, where there are potential threats due to unauthorized unmanned aerial vehicle (UAV) usage and intruders. These challenges include an inability to adapt to adverse situations, such as obstructions, interference, and dynamic changes in the environment, and a lack of large amounts of data for model training. Likewise, there are limits to the efficacy of models to generalize to new subjects or recognize new target classes (e.g., classes for which models have not been explicitly trained). One popular method is to utilize first principle physics modeling to augment data for training and estimation called Dynamic Data Driven Application Systems (DDDAS). Cognitive or fully-adaptive radar, has been proposed as a way of enabling radar systems to respond to dynamic environments. Currently, radio frequency (RF) transmissions are usually fixed and do not adapt to changing environmental conditions or target behavior. In contrast, cognitive radars optimize and adapt RF transmissions as part of a “Sense – Learn – Adapt” feedback loop – in response to dynamic changes in the environment. Although there has been much focus in the literature on waveform design and selection within a cognitive framework, there has been little attention to the resulting challenges to AI/ML approaches for ATR due to changing transmissions.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Apr 20, 2023
- Source ID
- FA95502210384
Entities
People
- Sevgi Zübeyde Gürbüz
Organizations
- Air Force Office of Scientific Research
- United States Air Force
- University of Alabama