Europe's satellite navigation network, protected by AI - predicting ground station failures before they happen

Galileo - Europe's own satellite navigation system - relies on a global network of ground stations called GESS (Galileo Experimental Sensor Stations). These stations continuously collect and transmit the data that keeps GPS-level positioning accurate for hundreds of millions of users across Europe and beyond. Historically, when something went wrong, engineers fixed it after the fact, or replaced parts on a rigid schedule regardless of actual condition. The stations were generating enormous volumes of telemetry - sensor readings, housekeeping data, signal quality metrics - every second of every day. None of it was being used to predict problems before they occurred. The technical challenge was substantial: very few historical outage records existed to train an AI on, the stations varied significantly in configuration, and no validated approach existed for applying modern machine learning to this class of space ground infrastructure.

E-Group (Prime Contractor) and Thales Alenia Space France (Subcontractor) developed two AI pipelines under ESA Contract 4000145430/24/NL/MH/mp, each designed to work around the data limitations of the real operational environment. The first pipeline uses classical machine learning - decision tree and gradient-boosting models - combined with a custom-built alarm-smoothing module called SmoothAlarm. SmoothAlarm solves a specific problem: raw AI predictions are noisy second-by-second outputs, which are operationally useless. SmoothAlarm accumulates predictions over a moving time window and only raises an alert when the evidence crosses a threshold calibrated against real historical outages - turning machine-level statistics into actionable maintenance alerts with meaningful lead time. The second pipeline uses TimesNet, a neural network architecture designed for time-series data with repeating patterns - the daily cycles, weekly rhythms, and maintenance routines that all ground stations follow. TimesNet learns what normal looks like, and flags behaviour that departs from it, without ever needing a list of past failures. It also identifies which specific signal channels are behaving unusually, giving engineers a starting point for investigation. A third direction - federated learning, which would allow multiple stations to collectively improve a shared AI model without sending their raw data to a central server - was also implemented and demonstrated as a prototype, pointing the way toward the operational-scale deployment.

All three project milestones were completed and the first two formally accepted by ESA; the Final Review was submitted in May 2026 and is pending formal acceptance. The classical ML pipeline with SmoothAlarm produces alarms with operationally meaningful lead times and positive-rate profiles - it tells engineers something is going wrong early enough to act. The TimesNet pipeline detects anomalies without requiring failure history, attributing unusual behaviour to specific signal channels so that investigation is targeted rather than blind. All experiments are tracked and reproducible in MLflow; the finalised AI model is registered and accessible to ESA. The project has demonstrated that applying modern AI to GNSS ground-segment maintenance is both technically viable and practically deployable. The natural next step - agreed at the Final Review with the ESA Technical Officer team - is to extend the validated approach from the 26-station experimental setup to the full operational Galileo ground segment, where several hundred stations generate the data volume needed to move from a feasibility result to a production-grade predictive maintenance capability.