Progressive Spatiotemporal Graph Modeling for Spacecraft Anomaly Detection.

Area of Science:

• Aerospace Engineering
• Data Science
• Artificial Intelligence

Background:

• Increasing numbers of spacecraft and telemetry data necessitate advanced anomaly detection.
• Existing methods struggle with spatiotemporal dependencies in multi-channel telemetry data.
• Accurate anomaly detection is vital for reliable spacecraft mission operations.

Purpose of the Study:

• To propose a novel Progressive Spatiotemporal Graph (PSTG) model for anomaly detection in multi-channel spacecraft telemetry.
• To overcome limitations of current methods in capturing complex inter-channel relationships.
• To enable accurate, real-time prediction and detection of anomalies in spacecraft telemetry.

Main Methods:

• Utilizing a multi-scale patch embedding module for hierarchical feature extraction and dimensionality reduction.
• Constructing a sparse adjacency matrix via multi-head attention integrating temporal, spatial, and cross-channel interactions.
• Employing an improved multi-head graph attention network to capture node dependencies.
• Incorporating a dynamic thresholding mechanism for online anomaly detection.

Main Results:

• The PSTG model effectively encodes rich spatiotemporal representations from intricate variable interactions.
• PSTG enables accurate, real-time prediction of multi-channel telemetry.
• Experiments on 84 months of real-world data show PSTG outperforms eleven state-of-the-art methods.
• Visualizations provide insights into spatiotemporal modeling and aid root cause analysis.

Conclusions:

• The proposed PSTG model offers a significant advancement in anomaly detection for multi-channel spacecraft telemetry.
• PSTG's ability to model complex spatiotemporal dependencies leads to superior performance.
• The model provides actionable insights for spacecraft operators, enhancing mission safety and efficiency.