Evaluating scalable similarity search and gaussian process regression for satellite telemetry anomaly detection on the ESA-ADB benchmark

The rapid growth of satellite constellations is pushing mission control toward automated Prognostics and Health Management (PHM), yet objective evaluation of anomaly detection methods has long been limited by the scarcity of labeled operational telemetry. Leveraging the European Space Agency’s anomaly detection benchmark (ESA-ADB), we (i) provide a quantitative characterization of spacecraft telemetry, summarizing key nominal challenges, and (ii) introduce an event-morphology taxonomy that groups annotated non-nominal intervals into eight archetypes. Building on these insights, we assess two detectors that answer the same telemetry constraints from opposite ends of the design space: a lightweight similarity search based on Sliding Minimum Euclidean Distance (SMED), and an enhanced Quantile Gaussian Process Regression (Q-GPR) with adaptive retraining and a spike-guard component. Under the full ESA-ADB hierarchical protocol on Mission 1, the two lead in different regimes: Q-GPR attains the highest event-wise F0.5 score on the lightweight channel subset (0.888 against 0.811 for SMED), whereas SMED leads on the complete set of 58 target channels (0.902 against 0.883). Both remain precise at scale, where the Telemanom-ESA-Pruned baseline collapses (F0.5 of 0.061 at a precision of 0.050 despite a recall of 0.870), and both improve on the out-of-limit detector representing the operational status quo (0.754 on the full set). What they require differs: SMED already scores 0.724 after nine weeks of history, against the 0.811 it reaches with the full record, and completes a full-set test pass in minutes rather than hours, whereas Q-GPR needs about three years of telemetry before it stabilizes but returns the better alarm quality on the lightweight subset and the better detection timing in both regimes. Being label-free and requiring neither per-channel model selection nor expert-driven retraining, both satisfy the constraints under which continuous monitoring of large channel populations is actually possible, where channels are downlinked periodically and require continuous processing.