Optimized Machine Learning-based strategies for on-board S/C failure detection: software integration and testing on a space-qualified processor
- Paper number
IAC-24,B6,IP,60,x85730
- Author
Dr. Antonio Leboffe, Thales Alenia Space Italia (TAS-I), Italy
- Coauthor
Mr. Davide DI Ienno, Thales Alenia Space Italia, Italy
- Coauthor
Ms. Ilaria Pinci, Thales Alenia Space Italia (TAS-I), Italy
- Coauthor
Mr. Francesco Corallo, Thales Alenia Space Italia, Italy
- Coauthor
Dr. Carlo Ciancarelli, Thales Alenia Space Italia, Italy
- Coauthor
Dr. Mauro Mangia, Alma Mater Studiorum - University of Bologna, Italy
- Coauthor
Ms. Livia Manovi, Alma Mater Studiorum - University of Bologna, Italy
- Coauthor
Dr. Alex Marchioni, Alma Mater Studiorum - University of Bologna, Italy
- Coauthor
Prof. Riccardo Rovatti, Alma Mater Studiorum - University of Bologna, Italy
- Coauthor
Ms. Eleonora Mariotti, Sapienza University of Rome, Italy
- Coauthor
Mr. Gianluca Furano, European Space Agency (ESA), The Netherlands
- Year
2024
- Abstract
Present-day spacecraft generate an enormous volume of telemetry data, continuously analysed for signs of potential system failures or anomalies. However, this reliance on human operators presents challenges, especially with the ever-increasing number of spacecraft in orbit. Automated and on-board anomaly identification and fault resolution become crucial for efficient and scalable operations. Complex on-board analyses are often impractical due to computational and power limitations. Furthermore, FDIR modules rely on pre-determined ''Out Of Limits'' checks for specific parameters (e.g., temperature exceeding a threshold). This approach lacks the ability to perform real-time multimodal analysis across multiple subsystems, potentially missing subtle anomalies or complex failure scenarios. Therefore, while FDIR systems offer valuable assistance by detecting and confirming pre-defined failures and triggering pre-programmed recovery actions, they fall short in addressing the evolving needs of complex spacecraft operations. Future advancements in onboard processing capabilities, combined with the development of AI-powered anomaly detection algorithms, hold promise for more sophisticated and adaptive fault management solutions, thus enabling robust and autonomous spacecraft operation. To overcome actual limitations, we propose to integrate Machine Learning (ML) algorithms to boost the FDIR functionalities along two directions, improve failure detection times and accuracy in the anomaly prediction. Various ML algorithms have been developed and tested using LEO EO Satellite's telemetry flight data. Three families of methods have been analysed: 1) Matching by Compression, 2) Matching by Prediction, and 3) Reference Models for anomaly detection. The first two families also include Deep Learning models: Convolutional Autoencoder for the former and LSTM-based recurrent neural network for the latter. The other most promising algorithms among these categories were: Principal Component Analysis and Autoregressive models. These algorithms have been trained on multivariate dataset created from the Reaction Wheel assembly telemetries and an assessment of their anomaly detection capabilities has been produced making a comparison with respect to the baseline approach. To cope with limited computational capabilities of GR740 space-qualified processor, the algorithms have been deployed into a self-contained standard C language library. Due to this, the library is highly portable and can virtually run on any platform that supports a C compiler. The performance of the algorithms was examined considering the most relevant metrics. The results include a benchmark for the GR740 board featuring a comparative assessment of the algorithms' performances under different conditions.
- Abstract document
- Manuscript document
IAC-24,B6,IP,60,x85730.pdf (🔒 authorized access only).
To get the manuscript, please contact IAF Secretariat.