Embedding Machine-Learned Anomaly Detection into Relevance-Driven Mission Operations Workflows

Anomaly detection has become a central application of machine learning in satellite operations; however, when applied in isolation, machine learning based detections alone are insufficient to support scalable mission operations. As fleets grow, operators face increasing cognitive load not from anomaly frequency, but from the need to interpret weak signals, correlate them with operational context, and decide whether and how to act. This paper presents an operationally grounded approach in which machine-learned anomaly scores are treated as inputs to a relevance-driven decision workflow rather than as standalone alerts.
We describe Mercury, an AI-based mission intelligence engine designed for continuous, mission-adaptive and unsupervised anomaly detection across telemetry streams, producing interpretable anomaly scores to characterize deviations from learned nominal behavior. These scores are not surfaced directly to operators. Instead, they are combined with additional signals from the ground segment—e.g. system state, operational modes, and scheduled activities—through a deterministic relevance-filtering layer. The system applies rule-based logic to combine anomaly probabilities with operational conditions, ensuring that only events with real operational impact are flagged. These operational events are enriched with contextual information required for investigation and decision-making.
The architecture is natively integrated with EASE-Rise, a cloud-native mission control and operations platform developed by Telespazio Germany. Telemetry and operational context flow from mission control to the intelligence layer, while prioritized events and recommended command sequences are fed back into the mission control environment, preserving established procedures and operator authority.
The approach has been evaluated on representative small satellite operational scenarios, where Mercury has demonstrated actual reductions in false escalations, investigation time, and routine monitoring effort, enabling operations to scale without linear increases in operator workload.