FIRST PRINCIPLES MACHINE LEARNING IN RADAR: AUGMENTING SIGNAL PROCESSING TECHNIQUES WITH MACHINE LEARNING FOR DETECTION, TRACKING, AND NAVIGATION

Abstract

Machine learning (ML) provides a set of tools for learning approximate system models from data. It has the potential to improve classic radar signal processing (RSP) algorithms by allowing them to maintain performance when the environmental assumptions used to derive them are violated. This could mitigate performance degradation experienced in more challenging scenarios, like those commonly found in airborne and maritime radar. However, the integration of ML into RSP algorithms presents a unique challenge due to the strict performance requirements of radar systems and often unpredictable nature of ML. This work examines an architectural approach to explainable ML that allows for the seamless integration of ML with more traditional algorithmic methods. This approach is then paired with causal ML concepts to develop a method for mitigating measurement drift in tracking and navigation. Next, an integrated system of low-cost ML systems are developed to enable adaptive detection algorithms to maintain CFAR-like performance across a range of interference distributions. Finally, generative ML techniques are used to reduce sample support requirements for adaptive detectors by directly constructing whitening filters from a small set of interference samples. This dissertation presents a framework for the successful integration of ML into RSP algorithms using a targeted approach based on a clear understanding of the first principles physics at play in a given application.