Abstract: Traditional satellite fault diagnosis methods and existing data-driven diagnosis methods both face challenges in identifying unknown faults that differ from known fault types,resulting in lower reliability and safety．To address the problem,a fault diagnosis and unknown fault detection method based on satellite digital twin and machine learning models is proposed．Firstly,various types of fault-simulated data are generated using satellite digital twin,and the fidelity of digital twin data are validated using XGBoost and real satellite fault samples, achieving the diagnosis of known fault types．On this basis,considering that existing methods cannot identify the occurrence of unknown fault types precisely,an out-of-distribution detection model Con-DAGMM is proposed,which is trained on normal data and known fault data to provide warnings for unknown fault．Experiments are conducted using digital twin data and satellite real fault data．The experimental results demonstrate that the proposed method achieves high fault diagnosis accuracy with an average accuracy of 98.8% on the test data．Furthermore,Con-DAGMM achieve high-performance unknown fault detection,outperforming Deep-SVDD and other comparison methods in precision,recall and F1 scores．The results indicate that satellite digital twin can overcome the scarcity of fault samples in satellite historical data,and the out-of-distribution detection approach can be successfully applied to warning of satellite unknown faults,enhancing the satellite's safety and reliability．

Key words: satellite control system, unknown fault detection, fault diagnosis, digital twin, machine learning, out-of-distribution detection