Research/ SPACE AI & AUTONOMY
SADDLE: Spacecraft Anomaly Detection using Deep Learning
Ankit Srivastava, Neeraj Badal, B. S. B. S. Manoj
Learning-based MethodsOpenAlex keyword: transformer models · Autonomy, Task Planning & OperationsOpenAlex topic: Anomaly Detection Techniques and Applications · Guidance, Dynamics & ControlOpenAlex topic: Fault Detection and Control Systems
Efficiently detecting anomalies in spacecraft data poses a significant challenge in modern space missions. We introduce Spacecraft Anomaly Detection using Deep Learning (SADDLE), a novel transformer network-based model tailored for spacecraft anomaly detection. SADDLE leverages an attention-based encoder to analyze telemetry data, capturing broader temporal trends critical for anomaly identification. It utilizes self-conditioning for robust feature extraction across multiple telemetry modalities. Additionally, SADDLE leverages Meta Gradient Descent to adapt faster to spacecraft data characteristics, simultaneously enabling effective training with limited anomaly examples. Extensive evaluations conducted on spacecraft datasets demonstrate that SADDLE outperforms all existing methods while significantly reducing training time.