A Generalized Software Framework for Deep Neural Network Inference in Space Applications

Recent advances in hardware and software technology have made it possible to implement more resourcedemanding deep learning algorithms in constrained hardware environments. This creates opportunities to use deep learning for aerospace applications on increasingly smaller spacecraft. This work shows our Neural Network Execution Framework, which aims to provide a cross-platform and reusable framework to deploy and execute trained neural networks for deep learning aerospace applications. This framework will execute the neural networks inference process regardless of the original deep learning framework in which they were created, for several flight software frameworks, operating systems, and hardware configurations. Starting with a specific implementation to execute one of our neural-based compression algorithms used to process images from the Solar Dynamics Observatory, using NASA's core Flight System as flight software, we recognized the potential to generalize our implementation to integrate more flight software frameworks and execute additional deep learning aerospace applications. Users and organizations can use this framework to create reusable deployment and execution solutions for deep learning, instead of implementing one-off solutions each time they need to develop a specific aerospace application. This approach allows developers to focus on their deep learning model architectures, rather than the implementation and deployment process on the flight platform. This work demonstrates the design, implementation, and testing of our framework, for deploying and executing neural networks developed in PyTorch and TensorFlow, including our compression algorithm, using the core Flight System and a low-cost single-board computer to emulate a space-like hardware-software configuration. This software demonstration aims to highlight the benefits of our framework for implementing deep learning in aerospace missions. Additionally, for comparative purposes, the performance of the networks in this hardware-software configuration is illustrated and contrasted with the performance obtained in a conventional computer environment.