Evidential Uncertainty-Aware Fault Diagnosis for Autonomous Spacecraft Health Management

Autonomous spacecraft must continue to make health-management decisions when communication with ground operators is delayed or temporarily unavailable. This creates a particular problem for fault diagnosis: a system must not only identify a known failure, but also recognize when the available telemetry does not provide enough evidence for a reliable diagnosis. Conventional rule-based Fault Detection, Isolation, and Recovery (FDIR) systems are generally conservative, while conventional machine-learning classifiers can remain highly confident when presented with failures that were not represented during training. This work develops an evidential diagnostic architecture for spacecraft health management. The proposed method represents competing fault hypotheses with a Dirichlet distribution and uses Mahalanobis distance on physically motivated telemetry features to determine the amount of evidence assigned to each hypothesis. Epistemic uncertainty is used to represent lack of knowledge about an observation, whereas aleatoric uncertainty captures ambiguity associated with noisy telemetry. The evaluation contains five controlled studies covering known-fault classification, uncertainty behavior, out-of-distribution (OOD) detection, telemetry-noise robustness, and component ablation. The benchmark includes 1,200 known-fault evaluations, 1,400 uncertainty-regime tests, 1,200 OOD frames, 2,100 telemetry-noise sweeps, and systematic component removals. On held-out known-fault frames, the proposed system obtains a Macro-F1 of 0.9533 with an Expected Calibration Error (ECE) of 0.0094. Under increasing telemetry noise, aleatoric uncertainty increases while epistemic uncertainty remains bounded for in-distribution observations. For previously unmodeled faults, epistemic uncertainty increases substantially, producing an OOD AUROC of 0.9422 and AUPRC of 0.9516. Maximum Softmax Probability (MSP), in contrast, obtains an AUROC of 0.4313. The experiments also expose two important limitations: cancellation in the metric representation during compound failures and reduced sensitivity to sensor-polarity reversals when magnitude-only features are used.