Research on Fault Diagnosis of Spacecraft Control System Based on Graph Neural Network

The fault diagnosis technology of spacecraft control systems is a crucial component of spacecraft health management. Traditional fault diagnosis technologies, such as model-based diagnosis and signal processing, are relatively mature but are limited in their ability to fully utilize information in big data environments. Neural network-based fault diagnosis methods often overlook the sequential nature of spacecraft data as time series data. Given that the state of a spacecraft at any given time is closely related to preceding data, it is possible to conduct fault diagnosis by extracting relevant features from fault sample data. In this paper, we propose a graph neural network-based fault diagnosis method for spacecraft control systems. To address the issue of limited fault samples, we construct a spacecraft control system model and generate various types of gyro and flywheel fault data through fault injection. Utilizing the Graphstar model—a powerful graph neural network with strong information mining capabilities—we transform one-dimensional fault data into graph data with spatial structure and extract relevant characteristics through deep learning to diagnose faults. Simulation results demonstrate that our method can effectively diagnose faults with high accuracy.