Daniel Aguilar Marsillach, University of Colorado Boulder; Marcus Holzinger, University of Colorado Boulder
Keywords: Autonomy, Raven-class Telescope Tasking, Evidential Reasoning, Custody Maintenance, Maneuver Detection
Abstract:
Given a limited amount of time and sensing resources, coupled with an increasing Space Object (SO) population, the SSA tactical sensor tasking problem is increasingly challenging. This problem is fundamentally a resource utilization one that determines how to use a finite collection of sensors to resolve specific hypotheses in as little time as possible. In the field of SSA, prior tasking methodologies aim to extract state estimates with minimum covariances, allowing operators to predict the dynamics of SOs with greater accuracy and reliability. Other approaches focus on implementing multi-target tracking with sensor tasking for improved space object catalog maintenance. These techniques are essential for accumulating data to resolve unknown-unknowns. However, limited work considers optimally using a suite of sensors to resolve known-unknowns, which links SSA hypotheses directly to sensor management.
Past work has been completed using Dempster-Shafer (DS) theory for sensor tasking and decision-making problems, which provides measures of certainty and ambiguity (ignorance). The quantification and qualification of such measures is important since decision-makers must act appropriately upon receiving such information. This motivated a general decision-making algorithm, known as Judicial Evidential Reasoning (JER), that optimally selects actions to reduce total system entropy. Using adversarial optimization, optimal sensor action sequences are computed that impartially resolve hypotheses by exploiting the principle of equal effort for adversarial agents. The algorithm is implemented in a receding time-horizon approach, where changing belief states provide the necessary feedback to determine the next set of actions sensors should take.
In this paper, JER is combined with estimation and reachability techniques to form a tasking algorithm that is able to evaluate and execute action sequences for optimal hypothesis resolution. In particular, the algorithm is updated to be purely driven by the system’s belief states. Two important aspects related to the maintenance of space situational and domain awareness are the ability to maintain SO custody and discern dynamical anomaly. Custody maintenance is important for real-time status monitoring while inferring dynamical anomalies is required for object re-acquisition and to conjunction event prevention. These two tasking objectives can be formulated in a frame of discernment using DS theory. To resolve these hypotheses, evidence for each has to be quantified.
This work redefines a Mahalanobis-distance binary detection method, specifically for Gaussian distributions, to quantify evidence for anomaly. Custody is quantified using position reachable set over-approximations and is maintained for a narrow-FOV telescope if the position reachable set approximation is contained in the telescope’s FOV. If the reachable set is not fully contained in the FOV, there exists a possibility that custody has been lost, favoring a search through the reachable set. The over-approximation is obtained by finding the maximal radius of the position reachable set and is updated for each SO in the observation campaign using continuation methods. The result is a telescope tasking algorithm that is driven to maintain custody by preventing position reachable sets from becoming too large while simultaneously detecting maneuvers.
A Monte Carlo study is designed and implemented to determine the accuracy and robustness of the algorithm to a specified statistical confidence level. SOs in the public SO catalog are modeled to perform aggressive North-South and East-West station maneuvers. With the use of confusion matrices, false-positive and false-negative classification rates are quantified. In terms of anomaly detection, it is found that reducing false-negative rates is most desirable for custody maintenance. Results show custody is maintained for all observable spacecraft in the Monte Carlo simulation.
In addition to presenting robustness in simulation, a successful and empirical real-time demonstration of the JER tasking algorithm is shown for a Raven-class telescope. This requires several working hardware and software subsystems. A telescope infinite-horizon LQR mount controller is used to open-loop track TLEs from the public catalog using TheSkyX Software. Image processing is used to background subtract, detect, track, and discriminate SOs of interest through the use of blob detection and a GM-PHD multi-target tracking algorithm. Topocentric right ascension and declination measurements are used with an Unscented Kalman Filter to estimate SO inertial states. The real-time demonstration resolves the custody and anomaly hypotheses for a combination of North-South and East-West station keeping GEO satellites.
This paper contains a number of contributions relating to the application and extension of the JER tasking algorithm for space situational awareness. First, a new maneuver detection method using mahalanobis distance is presented. Second, new evidence to hypothesis belief mappings are formulated that result in a purely belief and reachability driven algorithm. Third, the robustness of the algorithm is demonstrated through a Monte Carlo simulation. Lastly, the algorithm is interfaced with existing observatory hardware to resolve the listed hypotheses using optical data in real-time.
Date of Conference: September 15-18, 2020
Track: SSA/SDA