Multi-Modal Neural Network and Gaussian Process Models for Automated Space Domain Awareness

Toyon proposes multi-modal deep learning models for automated space object anomaly and maneuver detection, with the eventual development of an automated alert ontology for space operators in geosynchronous orbit and cislunar orbit. More specifically, Toyon proposes the development of deep kernel models, which leverage the expressivity and scalability of deep neural networks and the robust uncertainty quantification and data-centricity of Gaussian processes. Toyon will spend Phase I of the effort developing unimodal deep kernel models for anomaly and maneuver detection, with a particular emphasis on uncertainty quantification of model outputs. Toyon will leverage these algorithms to begin development of an end-to-end alert system prototype. In Phase II of the effort, Toyon will extend these unimodal models to the multimodal setting and perform robust analysis on the impact of data quantity and quality on model performance in a variety of space domain awareness applications.