Implementation of Lightweight Spacecraft Fault Diagnosis Network Based on FPGA Accelerator

• Published in: 2024 4th International Conference on Computer Communication and Artificial Intelligence (CCAI) pp. 497 - 502
• Main Authors: Lai, Jinlin, Wang, Siye, Nie, Saijun, Li, Yuandong, Mai, Ji
• Format: Conference Proceeding
• Language: English
• Published: IEEE 24.05.2024
• Subjects: Energy efficiency; Fault diagnosis; FPGA; hardware accelerator; Learning (artificial intelligence); lightweight; LSTM; Measurement; neural network; Pipelines; Power demand; Space vehicles; spacecraft anomaly diagnosis
• DOI: 10.1109/CCAI61966.2024.10603259

Exploring how to achieve a long-life, reliable, and stable on-orbit operation of spacecraft under limited hardware resource constraints is crucial. Artificial intelligence technologies are a promising solution for spacecraft fault diagnosis, thanks to their capabilities in data fusion and knowledge learning. This paper details the deployment of a highly efficient accelerator on the JFM7VX690T platform, developed by Fudan Microelectronics. Notable for its outstanding performance, the accelerator achieves precision and recall metrics of 98% or higher. Its efficiency is further enhanced by a multi-tiered pipeline architecture, which accelerates processing speeds and reduces latency. Additionally, an innovative encoding scheme eliminates the need for decoding, reducing computational burdens. Operating at 135.2 Giga Operations Per Second (GOPS) while consuming only 1.138 watts of power, the accelerator is markedly more energy-efficient and has a significantly lower power consumption compared to traditional Central Processing Units (CPUs) and Graphics Processing Units (GPUs).