Dwight Temple, ExoAnalytic Solutions
Keywords: neural, network, anomaly, characterization, SSA, autonomous, event, classification, indication, warning
Abstract:
Neural networks, specifically convolutional neural networks (CNNs) and generative adversarial networks (GANs), have permeated seemingly every scientific field. Automated vehicle navigation, medical scan anomaly detection, adversarial network attack/prevention, unknown state-space tracking, and recently space situational awareness have shown provocative results for some of the most arduous applications. An industry standard is to utilize CNNs for supervised learning approaches for image classification and to use GANs for unsupervised learning where replication is the desired outcome. Proposed in this paper is a fusion of these methods to be operated in series to produce life-like representations of spacecraft anomalies in order to discriminate successfully these events from non-anomalous operating spacecraft behaviors. ExoAnalytic Solutions, with over 225 autonomous ground-based sensors, collects over 500,000 spacecraft images nightly; this vast database serves as an increasingly powerful resource bank that drives innovation for robust space-situational awareness (SSA) and neural network advancements. With recent debris-generating events occurring at geostationary orbit (GEO), proactive monitoring of spacecraft could permit avoidance of these disastrous events through specific indications and warnings (I&W). Specific events of interest include close approaches, glinting, plume generation, breakups and other debris generating events. Discerning between nominal and anomalous operation with limited metrics, elevated noise, and star obfuscated focal planes further complicates the problem. The hybrid architectures used and explored in this paper utilize disparate complete architectures as its components. For example, several GANs will be used for individual generative tasks, while stacked CNNs are explored for data segmentation. Data utilized from ExoAnalytic Solutions will power a network of networks specifically operating on non-resolved satellite imagery. This paper describes the strengths of various neural networks for SSA applications and applies a set of CNNs and GANs to an example of a relevant application, focusing on the discrimination of close-approach maneuvering and nominal operation utilizing collected and synthetically generated data.
Date of Conference: September 11-14, 2018
Track: Poster