Toward an Onboard Anomaly Detection and Identification Method for Satellites

Unexpected anomalies can occur when satellites are exposed to harsh environments. In addition, operators cannot directly repair or inspect satellites, making stable operations reliant on the early detection and identification of anomalies. However, conventional data-driven satellite monitoring methods often lack explainability and feasibility for onboard implementation. This study proposes a novel onboard satellite monitoring method that detects anomalies, identifies their types, and locates specific abnormal channels within housekeeping data. The proposed method employs the Mahalanobis distance for anomaly detection owing to its low computational cost, making it suitable for onboard applications. Feature contribution analysis using orthogonal arrays and signal-to-noise ratios was used to enhance explainability, attempting to identify the type and location of detected anomalies. Through experiments using benchmark datasets, we demonstrated that the proposed method achieves detection performance comparable to that of conventional machine learning-based methods expected to be implemented at ground stations, and can identify anomaly locations and anomaly types to a limited extent. In addition, execution speed tests conducted on a Raspberry Pi Zero confirmed its feasibility for onboard implementation. These results suggest that the proposed method contributes to the realization of a self-diagnostic capability for satellites, thereby enhancing their reliability and operational stability.