Probabilistic graph neural inference aims to improve satellite anomaly response with built-in auditability

A developer working on satellite anomaly triage found that a temporal autoencoder and threshold rules missed a reaction wheel fault that appeared as correlated, sub-threshold drift across several subsystems. Seeking a fix, they explored graph neural networks, where subsystems are nodes and their physical and functional dependencies are edges, so each subsystem's score can reflect its neighbors' states. They argue that operators need calibrated probability distributions over failure modes, not binary verdicts, and that the reasoning behind each output must be traceable for later review. The write-up walks through the graph construction, probabilistic inference method, audit trail design, and lessons from a prototype tested on simulated data.