A greedy worst-case framework for online, interpretable anomaly detection in spacecraft telemetry

Abstract

Spacecraft telemetry anomaly detection must operate under stringent operational constraints: physically heterogeneous channels, rare unlabeled faults, a nominal regime contaminated by sensor glitches and drifting across mission phases, resource-limited online execution, and a requirement to explain every alarm. Simple per-limit thresholding is brittle and masked by contamination; deep sequence models are accurate but heavy, opaque, and their point-adjusted scores are known to inflate performance. We present a greedy worst-case detector that is lightweight, interpretable, and calibrationfree. Channels are placed on a common footing by a robust MAD-consistent standardization (50% breakdown point); a single-step greedy rule selects the worst-case channel per step—a safety/minimax choice; and the resulting max statistic is judged by an extreme-value/Šidák threshold calibrated to the effective number of correlated channels meff, so that detection requires no labelled anomalies. Flagged channels are ranked as candidate root causes, giving interpretability by construction. The method is single-pass O(m) . On the public NASA SMAP/MSL benchmark (81 real spacecraft channels), under a matched fixed-alarm-rate protocol and with AUC reported as a threshold-free ranking, the calibrationfree Šidák( meff) threshold attains the highest macro-F1 , F0.5 and recall of any method, while deep reconstruction models—including a trained Anomaly Transformer—lead on threshold-free AUC where temporal context is decisive; the detector runs roughly three orders of magnitude faster per sample. An equicorrelated-Gaussian family-wise-error-rate study and an empirical false-alarm check on the benchmark’s nominal telemetry confirm the threshold calibration, and a qualitative case study illustrates the ranked attribution. Point-adjusted F1 is reported alongside raw metrics throughout.