Multiforecast-based Early Anomaly Detection for Spacecraft Health Monitoring
Prajjwal Yash, Sharvari Gundawar, Nitish Kumar, B. R. Uma, Krishna Priya Ganesan, Purushottam Kar · 2024
Early detection of impending anomalies is a strong desirable for spacecraft operation as it can allow preemptive action to safeguard the mission objectives. Methods abound for just-in-time anomaly detection but early detection is a much more sought after goal. In this paper, we present MEND, a simple-yet-powerful model for early anomaly detection for spacecraft health monitoring. In experiments, MEND was able to provide strong alerts for impending anomalies as much as 10-15 minutes before the onset of the anomaly which could give a system admin valuable time to perform curative action. It is notable that none of the other models considered, including state-of-the-art zero-shot time series prediction models, were able to achieve this. MEND is based on simple, explainable elements such as self-supervised operation-mode detection and self-disagreement-based anomaly detection which, as a side effect, offer insights that may aid root cause analysis and may be of independent interest. Code for MEND is available at https://github.com/purushottamkar/mend