From Event Logs to Action: Procedure Grounded Anomaly Prioritization for Human Machine Teaming in Satellite Operations
- Paper ID
111872
- authors
- company
Korea Institute of Energy Technology (KENTECH); ; Korea Aerospace Research Institute (KARI)
- country
Republic of Korea
- year
2026
- abstract
Satellite operations teams face continuous streams of event logs during ground contacts, but only limited time to review each pass. Missed or delayed identification of abnormal behavior increases response time and operational risk, while sparse and heterogeneous fault labels make fully supervised approaches impractical. We propose an approach that ranks contact windows for review and, for each flagged window, provides specific event evidence and relevant operations procedures for operator confirmation. To do so, we model each pass window as an event-transition graph and learn a nominal reference transition structure directly from unlabeled logs to support human-machine teaming (HMT) in operational decision-making. We define contact windows by scheduled pass contacts from Acquisition of Signal (AOS) to Loss of Signal (LOS) and represent each window as a directed transition graph over event types. An attention-based graph neural model is trained in a self-supervised manner via link prediction to learn transition patterns that are most commonly observed during routine operations. Each observed transition receives a risk score based on how unlikely it is under the learned reference transition model. Window-level risk then combines repetition amplification, which surfaces abnormal repeats, with exponentially decayed accumulation, which captures sustained abnormality. For operator-facing review, the method surfaces the most anomalous subsequences and repetition statistics, and provides ranked diagnostic hypotheses and candidate recovery procedures through a lightweight mapping between event signatures and validated operations procedures. We evaluate the approach on a year of real-world spacecraft operations artifacts with sparse incident labels and show that it ranks fault-related windows higher in the ranked review list than Markov transition-probability and Petri net baselines. Under a fixed review budget, operators can focus on a small set of high-risk windows while still capturing a substantial portion of fault-related cases. Case studies show that the method surfaces abort-retry motifs, such as repeated playback start/stop, and points operators to relevant procedures, supporting faster operator assessment.