Dan Shen, Intelligent Fusion Technology, Inc; Carolyn Sheaff, Air Force Research Laboratory (AFRL); Genshe Chen, Intelligent Fusion Technology, Inc; Mengqing Guo, Intelligent Fusion Technology, Inc; Erk Blasch, Air Force Office of Scientific Research (AFOSR); Khanh Pham, Air Force Research Laboratory (AFRL)
Keywords: Machine Learning, SSA, SDA, Game Theory, Generative Adversarial Network, Persistence of Excitation, Maneuver Detection and Classification
Abstract:
Space superiority requires space protection and space domain awareness (SDA), which rely on rapid and accurate space object behavioral and operational intent discovery. The satellite maneuver detection and classification is the first step of space behavior discovery. With exiting capabilities based on anomaly detection, classifying orbits as normal or abnormal ones is lacking maneuver details to support space object behavioral and operational intent discovery. Consequently, an automatic maneuver detector with enough details for SDA in the detection results is desired.
Machine learning and artificial intelligence (ML/AI) is good candidate for satellite maneuver detection and classification. The generative adversarial networks (GANs) are data-hungry and rely heavily on vast quantities of diverse and high-quality training examples in order to generate accurate satellite maneuver detectors. However, unlike image processing, the SDA has limited training data available. Moreover, GANs remain remarkably dif?cult to train and existing approaches to improve GANs train data efficiency still rely on heuristics that are extremely sensitive to modi?cations.
In this paper, we modify the game setup of generators and discriminators in GANs. The zero-sum game in existing GANs is replaced by a general sum game with objective functions containing small Gaussian disturbances. We will train GANs with a Fictitious play concept framework, within which each player (generator or discriminator) presumes that the opponents (discriminator or generator) are playing stationary (possibly mixed) strategies. At each round, each player thus best responds to the empirical frequency of play of their opponent. The GAN contains two adaptive control problems. In the adaptive control of the generator, the generator tries to obtain the discriminators control policy then optimize his own objective function. The same strategy works for the discriminator. From the perspective of adaptive controls, these small Gaussian noises in the cost function can help satisfy the persistence of excitation, which guarantees convergence without a priori stability assumptions and ensures robustness properties. Thus, the proposed modification improves the training data efficiency and enables the application of GANs in the SDA.
The modified GANs are applied in the satellite maneuver detection and classification from the ground-based sensing data, which includes the Azimuth angle (rad), elevation angle (rad), range (km), range rate (km/s), principal RCS (m^2), and orthogonal RCS(m^2) of space objects. We simulate the low Earth orbits (LEOs) based on the two-line elements (TLEs) from space-track.org and the simplified general perturbation version 4 (SGP4) propagator. A small set of 100 orbits and 100-second simulated tracks with various maneuvers (different pointing angles and magnitudes, and maneuvering periods) are used for training the proposed GANs. We used resampling and interpolation to normalize the input data to the GANs. The sampling frequency of the input data is 0.08 Hz, ie, the gap between the sensing data is 12.4 seconds.
The modified GANs demonstrate the training convergence on the small set of training data. We achieved 94.5% accuracy on the evaluation data. To evaluate the proposed GANs for SDA, we also compute several quantitative measures, such as average log-likelihood, inception score (IS), Frechet Inception Distance (FID), maximum mean discrepancy. We obtained 9.74 in IS and 0.002 in FID. These performance metrics show the proposed GANs for SDA can automatically, accurately, and rapidly detect and classify the satellite maneuvers to support the further space behavioral and operational intent discovery.
Date of Conference: September 27-20, 2022
Track: Machine Learning for SSA Applications