Mode Selection in Cognitive Radar Networks

TL;DR

The paper addresses energy-efficient mode selection in cognitive radar networks by allowing nodes to switch between active radar and passive ESM based on learned target behavior. It introduces a two-tier modeling framework where targets are represented by dual Markov processes (motion and emissions) and organized into classes to guide mode choices. Centralized and distributed bandit-based algorithms are proposed, with class formation via Wasserstein-clustering and MT-PHD tracking to sustain accurate state estimation. Numerical results demonstrate that class-informed strategies can reduce radiated power while maintaining or improving tracking accuracy, though network latency can affect centralized performance. The work provides a foundation for scalable, adaptive CRN operation with practical implications for SWaP-constrained sensing and surveillance systems.

Abstract

Cognitive Radar Networks, which were popularized by Simon Haykin in 2006, have been proposed to address limitations with legacy radar installations. These limitations include large physical size, power consumption, fixed operating parameters, and single point vulnerabilities. Cognitive radar solves part of this problem through adaptability, using biologically inspired techniques to observe the environment and adjust operation accordingly. Cognitive radar networks (CRNs) extend the capabilities of cognitive radar spatially, providing the opportunity to observe targets from multiple angles to mitigate stealth effects; distribute resources over space and in time; obtain better tracking performance; and gain more information from a scene. Often, problems of cognition in CRNs are viewed through the lens of iterative learning problems - one or multiple cognitive processes are implemented in the network, where each process first observes the environment, then selects operating parameters (from discrete or continuous options) using the history of observations and previous rewards, then repeats the cycle. Further, cognitive radar networks often are modeled with a flexible architecture and wide-bandwidth front-ends, enabling the addition of electronic support measures such as passive signal estimation. In this work we consider questions of the form "How should a cognitive radar network choose when to observe targets?" and "How can a cognitive radar network reduce the amount of energy it uses?". We implement tools from the multi-armed bandit and age of information literature to select modes for the network, choosing either an active radar mode or a passive signal estimation mode. We show that through the use of target classes, the network can determine how often each target should be observed to optimize tracking performance.

Figures (11)

• Figure 1: Maximum intercept range for a single radar node with a variable transmit probability.
• Figure 2: A model of the type of network we consider, where each node can choose between active and passive observation of several types of targets. The nodes send observations to the fusion center.
• Figure 3: ESM receiver instantaneous SNR for a center frequency of $\qty{1}{\giga\hertz}$, omnidirectional receive antennas (gain of $\qty{0}{dB}$), and $\textcolor{black}{\qty{3}{dB}}$ of losses. Given an SNR requirement of $\qty{0}{dB}$ for detection 69145495757489, targets can be detected at ranges of up to $\qty{100}{\kilo\meter}$, depending on the transmitter power.
• Figure 4: ESM receiver instantaneous SNR for multiple transmit antennas and orientations. The same system parameters as Figure \ref{['fig:ESM_SNR']}, with a 1 Watt transmit power.
• Figure 5: Kalman filters which are tuned to the process noise and motion model probabilities for the class will provide lower tracking error than equivalent filters which are "untuned".

Theorems & Definitions (6)

• Definition 1: Equal in State Distribution
• Definition 2: Target Class
• Definition 3: Target Family
• Proposition 1: Unique Class
• proof : Proof of Prop. \ref{['prop:unique']}
• Definition 4: Maximum Detectable ESM Range