An interpretable graph neural network for real-world satellite power system anomaly detection based on graph filtering

• Published in: Expert systems with applications Vol. 254; p. 124348
• Main Authors: Di, Yi, Wang, Fujin, Zhao, Zhibin, Zhai, Zhi, Chen, Xuefeng
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 15.11.2024
• Subjects: Anomaly detection; Graph neural network; Interpretability; Satellite power system; Satellite power system; Interpretability; Graph neural network; Anomaly detection
• ISSN: 0957-4174
• DOI: 10.1016/j.eswa.2024.124348

Detecting anomalies in satellites holds immense importance within the aerospace industry. Many current detection methods only focus on temporal correlations and ignore spatial correlations. To take into account both temporal and spatial features, we proposed to employ Graph Neural Networks (GNN). In fact, to develop a GNN approach that is applicable to the satellite power system anomaly detection in real-world scenarios, it is imperative to take into account the specific requirements of the aerospace industry. Our analysis and research indicate that the critical factors lie in model interpretability and data characteristics. First, as model interpretability can bring trustworthiness, traceability, and guidance, it is indispensable for the application of deep learning algorithms in satellite telemetry data anomaly detection. We created a novel graph filter: Adaptive Quadratic Approximation Graph Filter (AQAGF), which is capable of both approximation ability and response time. It is transformed into a spectral graph convolution neural network that can draw the frequency response function to provide the model interpretability. Then, to address the difficulties of telemetry data characteristics: strong noise and missing data, we introduced autoencoder architecture and adversarial training strategy to further design our anomaly detection method, called Interpretable Spatial-Temporal Graph Anomaly Detection (ISTGAD) based on AQAGF, which is robust to noise and data missing. And it is able to capture temporal and spatial correlations together within a transparent working mechanism. Finally, experiments proved that our model has better anomaly detection performance than other state-of-the-art methods, and we also provide visualizations to showcase the interpretability and working mechanism of our model. Our code and data are publicly available at: https://github.com/DiYi1999/ISTGAD.