ESA GNC Conference Papers Repository

Smart-FDIR: An Unsupervised Deep Learning Approach to On-Board AOCS FDI(R), First Results and Conclusions
Arthur Scharf, Carlos Hervas Garcia, Calum Murray, Dave Thomas
Presented at:
Virtual Conference 2021
Full paper:

The Fault Detection, Isolation and Recovery (FDIR) system on any modern spacecraft is a critical component with respect to the performance, safety, autonomy and availability of any space mission. Today?s FDIR solutions rely on complex and detailed failure analyses and require laborious design and tuning of dedicated monitors. Despite this complexity, FDIR solutions are often good at detecting system failures but rather limited in their isolation capabilities. SMART-FDIR provides a novel approach to Failure Detection and Isolation (FDI) leveraging deep learning techniques to reduce development cost and complexity of the overall on-board FDIR function. The AI model is trained on high-fidelity simulator data to distinguish between nominal and anomalous behaviour, without a-priori knowledge of anomaly signatures. This paper gives a summary of state-of-the-art anomaly detection techniques and discusses the solution chosen for SMART-FDIR. It then provides results on the selected approach for multiple use cases including real telemetry data from Solar Orbiter, S5P and SMAP, using the novel deep learning paradigm named MODISAN - Modificator-Discriminator Adversarial Networks -, the backbone algorithm of SMART-FDIR. A benchmark and direct comparison between classical FDIR techniques and SMART-FDIR is presented, as well as an overview of the implications of such an approach on established Functional Avionics Processes.