Interpretable Anomaly Detection of Multisatellite Attitude Control Systems Based on TabGMAN

The health status of multisatellite attitude control systems exerts a crucial impact on the operational efficiency of constellations. However, the inherent spatiotemporal correlation and heterogeneity of multisatellite telemetry data increase the anomaly detection difficulty using traditional methods. To address the challenges in multisatellite attitude control system health management, this article proposes an interpretable anomaly detection method based on the tabular graph multiattention network (TabGMAN). First, this article integrates constellation orbital and component information and time–frequency domain features to construct multisatellite graph data. Subsequently, we propose the TabGMAN, an interpretable spatiotemporal neural network that adopts a spatiotemporal attention mechanism and instancewise feature selection to identify anomalous satellites and attitude control system components. Finally, in several cases of anomaly detection in constellations, the proposed method achieves effective identification of anomalous satellites and attitude control system components and fault attribution through quantitative interpretability analysis.