Rasit Abay, UNSW Canberra; Steve Gehly, UNSW Canberra; S. Balage, UNSW Canberra; Melrose Brown, UNSW Canberra; Russell Boyce, UNSW Canberra
Keywords: Machine learning,maneuver detection, generative adversarial networks, space situational awareness
Abstract:
It is essential to detect the maneuvers conducted by resident space objects that don’t have readily available telemetry to promote space situational awareness. Maneuver detection and characterization in near-real time are helpful for identifying anomalous orbital behavior and associated threats, such as close approaches, to the other neighboring space assets. In addition, such capability facilitates the maintenance of space object catalogs in terms of reducing the workload that is necessary for recapturing the lost objects between observations. Although there are studies that investigate the feasibility of detecting and characterizing maneuvers from observations, they are compute-intensive. This paper presents a deep generative model that can learn the nominal orbital behavior of resident space objects and detect maneuvers in near-real time from orbital evolution.
The input data are simulated orbit dynamics with high fidelity force and realistic space environment modeling to represent the data distribution that is similar to real observations. Because the success of machine learning methods depends on the representation of the input data, the simulated orbital trajectories are represented in various orbital elements to find the most appropriate form that improves the learning of the neural network. In addition, data augmentation techniques are used to augment the learning rate of the model. The simulated orbital data are used to train, evaluate, and test the performance of the neural network architecture. During training, two neural networks, namely generative and discriminative models, are trained to learn the data distribution and to estimate the probability that the given sample data whether come from the training data respectively. When the minimax two-player game framework between generative and discriminative models is successfully finalized, the generative model is able to generate the training data distribution and discriminative model can’t distinguish it from the training data.
This paper investigates the feasibility of using generative adversarial networks to detect the maneuvers conducted by space objects. Because the generative model can recover the high dimensional and nonlinear training data distribution to represent the expected normal orbital behavior, any unexpected orbital evolution can be detected. Although machine learning models are trained with high fidelity force modeling, real data can be used to improve the accuracy of the models by applying transfer learning.
Date of Conference: September 11-14, 2018
Track: Space Situational Awareness