STGLR: A Spacecraft Anomaly Detection Method Based on Spatio-Temporal Graph Learning.

Area of Science:

• Aerospace Engineering
• Data Science
• Machine Learning

Background:

• Spacecraft operations are susceptible to anomalies, necessitating robust detection methods.
• High-dimensional, complex, and large-scale telemetry data pose challenges for existing anomaly detection techniques.
• Current methods often overlook inter-variable correlations and struggle with inferring initial variable relationships due to a lack of prior knowledge.

Purpose of the Study:

• To propose a novel spatio-temporal graph learning reconstruction (STGLR) method for effective spacecraft anomaly detection.
• To address the limitations of existing methods in capturing complex correlations within spacecraft telemetry data.

Main Methods:

• Employs dynamic graph learning to infer initial relationships among telemetry variables.
• Utilizes a spatio-temporal feature extraction module with a graph sample and aggregation network for dependency analysis.
• Incorporates an attention mechanism for adaptive feature selection and a reconstruction module for pattern learning.

Main Results:

• The STGLR method demonstrates superior performance in spacecraft anomaly detection compared to existing approaches.
• Experiments on two public spacecraft datasets show STGLR's effectiveness in capturing normal telemetry patterns.
• Achieved an average F1 score exceeding 0.97, indicating high accuracy in anomaly identification.

Conclusions:

• The proposed STGLR method offers a significant advancement in spacecraft anomaly detection.
• Its ability to learn complex spatio-temporal dependencies and variable correlations enhances detection accuracy.
• STGLR provides a reliable solution for ensuring the normal operation of spacecraft through advanced anomaly detection.