SBIR/STTR Award attributes
As hypersonic technologies emerge, key challenges related to navigation must be addressed. Size and weight requirements impose limitations on both sensor quality and processing power, Global Navigation Satellite Systems (GNSS) positioning is known to be unreliable in contested environments, and classical approaches to error dynamics characterization breaks in high-acceleration conditions. In addition, numerical errors imposed by linearization and Gaussian assumptions compound the issue during dead reckoning of inertial navigation systems (INS). Most of the error processes behind traditional INS error dynamics are fundamentally driven by statistical random noise which imposes a theoretical upper bound on the performance of any INS. The path to approaching the upper bound lies in perfectly characterizing all the nonlinear relationships and eliminating sources of numerical error that result from assumptions that are liable to break. This list includes accounting for inertial measurement Unit (IMU) bias stability, scale factor errors, axis misalignments, axis nonorthogonalities, finite sample rates, gravity errors, to name a few. Given the ability for deep neural networks to learn complicated nonlinear functional relationships, EpiSci proposes SHEPERD to address the problem of error prediction and error correction for navigation in hypersonic applications. The proposed SHEPERD is a modular architecture that leverages machine learning transformers and is designed to address the above challenges while also addressing concerns related to the future transition from simulation to real platform hardware. Phase I will include a study of the architectural trade space, an implementation of the proposed architecture, and an evaluation of the proof-of-concept software. Phase II will focus on expanding the SHEPERD capability and robustness via increased model fidelity and establishment of subsystem requirements driven by anticipated hypersonic platform limitations and requirements. Approved for Public Release | 22-MDA-11215 (27 Jul 22)
