A versatile feature selection learning method for satellite attitude control system fault diagnosis with limited data

Fault diagnosis in satellite attitude control systems is crucial for ensuring the operation and maintenance of satellites. However, current methods face challenges in achieving the desired classification results due to limited fault samples and complex temporal characteristics. To address these problems, this paper proposes a versatile fault diagnosis method called feature selection learning, which divides attitude control system faults into component-level and system-level to create a fault hierarchy. The method considers an overfitted classifier as the base-classifier (BC). It then sifts through a subset of downscaled fusion features derived from the BC to serve as inputs for the feature selection embedder (FSE), aiming to identify the FSE that captures the core fault features. The efficacy of the proposed method in each group of fault diagnosis is confirmed using the satellite attitude control system dataset. The experimental results show that the proposed method achieves a fault diagnosis accuracy of 91.08% under extremely scarce sample conditions, which is an increase of 4.69% compared to the optimal accuracy among the five comparison methods. Furthermore, the proposed method exhibits stable fault diagnosis performance across different datasets, sample sizes, and BCs, demonstrating its applicability in the context of fault hierarchy.