Capuano, Giovanni Maria (2025) Edge AI for Autonomous Spacecraft: Enabling Onboard Intelligence with FPGA Acceleration. [Tesi di dottorato]

Documento PDF capuano_giovanni_maria_38.pdf Visibile a [TBR] Amministratori dell'archivio Download (130MB) | Richiedi una copia

Abstract

Artificial intelligence (AI)-driven onboard data processing is a cornerstone of future space missions, enabling autonomous and time-critical decision-making in applications such as planetary exploration, in-orbit servicing, and Earth Observation (EO). Executing AI algorithms directly onboard mitigates the latency and bandwidth constraints inherent to ground-based processing. However, most commercial off-the-shelf edge-AI accelerators lack the radiation tolerance required for long-duration and deep-space missions. Space-qualified Field-Programmable Gate Arrays (FPGAs) provide a robust alternative, combining fault tolerance, low power consumption, and in-flight reconfigurability. This thesis investigates FPGA-based acceleration of deep learning models for real-time target detection in optical Earth observation. The proposed SR-YOLOv5s network integrates a convolutional super-resolution backbone to enhance the recognition of small maritime targets, achieving a mean mAP50 of 0.9272 ± 0.0029 and a 64.5% gain in very small object detection. Implemented on the Microchip PolarFire System-on-Chip with the CoreVectorBlox accelerator, the model achieves 18 fps while consuming less than 2 W of power, demonstrating the feasibility of compact, energy-efficient neural pipelines for size-, weight-, and power-constrained platforms such as CubeSats. Integration of the CoreVectorBlox within the ANHEO autonomous navigation unit demonstrated the advantages of FPGA reconfigurability through bitstream updates, yielding a 3.3× performance improvement and a 50% energy reduction per frame in the satellite detection task. Altogether, the findings establish reconfigurable FPGAs as a key enabler of intelligent and resilient spacecraft, providing the capability to adapt onboard computational resources to evolving mission objectives, operational conditions, and algorithmic developments.

Downloads

Actions (login required)

Modifica documento