A causal graph-based framework for satellite health monitoring

In satellite operations, one of the essential tasks is to monitor the health status of the systems, which involves forecasting telemetry data that reflects the state of health. The application of data-driven approaches in system monitoring has led to significant improvements in health monitoring and anomaly detection. However, existing methods fail to fully leverage the complex intersensor relationships present in satellites. They do not explicitly exploit the structure of these relationships to predict the expected behavior of telemetry time series either. To address these limitations, this paper introduces a novel health monitoring framework for artificial satellites that combines causal graphs and deep learning. In the causality learning phase, we propose a method that integrates mRMR (Maximum Relevance Minimum Redundancy) and PCMCI (Peter-Clark Momentary Conditional Independence) to construct an efficient and accurate causal discovery approach for learning causal graphs for high dimensional telemetry data. Subsequently, we design a graph attention-based neural network that incorporates these causal graphs into a deep network for prediction. Experimental evaluation on two datasets from satellite attitude control systems and power systems demonstrates the superior performance of our proposed method in accurately predicting health status compared to baseline approaches. Furthermore, the experiments highlight the interpretability-enhancing role of causal graphs, which is beneficial for health monitoring and anomaly detection.