Improving AI Monitoring of Early Life Satellites Using Transfer Learning

Résumé

In the last decades, many space domain actors such as the Centre National d’Etudes Spatiales (CNES) have begun to use Artificial Intelligence to monitor spacecraft housekeeping telemetry. These novel techniques are able to identify atypical behaviours and potential satellite anomalies that cannot be detected by more standard monitoring approaches. However, AI methods have an important drawback: they need a significant amount of data to be able to “learn” the nominal behaviour of a spacecraft and then detect novelties in new telemetry, which is not suitable for a satellite in the beginning of life where in-flight telemetry is very scarce. One way to bypass the scarcity of data is Transfer Learning (TL). Depending on the use case, operators may have already-available telemetry either from on-the-ground Assembly, Integration, and Test (AIT) of the spacecraft, from full-digital or hybrid simulators, or from in-flight telemetry of one or multiple “twin-spacecraft” in case of a constellation with already-launched units. This already-available telemetry is often close, but not perfectly similar, to in-flight telemetry of the newly-launched spacecraft to be monitored. The idea of TL is therefore to use this large and existing database (the source database), coupled with the first in-flight telemetry from the new spacecraft (the target database), to be able to mathematically-design a relevant AI learning model. In 2022, CNES and TéSA laboratory have worked together and have identified two TL methods to detect anomalies in telemetry of early life satellites with few data, by working directly on the telemetry dataset (the learning domain) or on the model learned from the target database. The first TL method consists in mathematically modifying the decision boundary estimated by a One-Class Support Vector Machine (OC-SVM) algorithm applied to the source database to match the target database. The second method based on “Domain Transfer” consists in building an “extended” learning domain made up with the relevant data from both the source and target databases, which is used to build a learning model. These two algorithms have been evaluated with real Earth Observation satellite telemetry. The preliminary outcomes of this research show promising results. Further work will consist in implementing these methods operationally so that AI monitoring methods can be used from the very beginning-of-life of CNES satellites. The main conclusion of this work is that TL can be an interesting tool to monitor spacecraft housekeeping telemetry during the first 6 months after the launch of a satellite.