Implementing a Neural Network Execution Framework in Realistic Space Hardware and Software as a Pseudo On-Orbit Demonstration

Implementing a Neural Network Execution Framework in Realistic Space Hardware and Software as a Pseudo On-Orbit Demonstration

Utah State University, Logan, UT

Recent advances in hardware and software technology have made it possible to implement more resource-demanding deep learning algorithms in lighter hardware environments. This creates opportunities to use deep learning for space applications on increasingly lighter and smaller spacecraft. The goal of this work is to demonstrate the viability of implementing a Neural Network Execution Framework (NNEF) that can facilitate a cross-platform and unified deployment of any neural network onboard a spacecraft hardware and flight software. The NNEF generalizes the neural network inference process, regardless of the original framework in which they were created. This allows users to focus on the development of their scientific model architecture and deep learning objectives, rather than being distracted by the implementation process onboard the spacecraft. This framework has been implemented to run inside NASA's core Flight System and on top of a Raspberry Pi 4 board, demonstrating the capability to execute a variety of trained neural networks created in Pytorch and Tensor Flow. This includes a neural-based compression algorithm used to process images from NASA's Solar Dynamics Observatory in a space-like hardware-software configuration. This initial software implementation shows the feasibility of our goal, demonstrating the deployment of deep learning benefits through our framework in a unified way for a broader range of space missions and applications. In addition, for comparison purposes (not for benchmarking), it showed the performance of the networks running in the mentioned hardware-software configuration contrasted with the performance obtained in a regular computer environment.