Abstract: With the development of aerospace science and technology, intelligent fault diagnosis technology is one of the key technologies to ensure safe and autonomous operation of spacecraft control system. Due to the small number of unlabeled telemetry data samples with high noise, it is difficult to diagnose fault signals accurately of the spacecraft in orbit by traditional fault diagnosis methods. A deep transfer learningbased fault diagnosis method is proposed to realize realtime fault diagnosis of spacecraft in orbit. First, onedimensional timedomain signals are converted into twodimensional image signals to realize the preprocessing of spacecraft operation dataset. Secondly, a residual networkbased deep learning fault diagnosis framework is built and pretrained via ground trained dataset and onorbit operation data of other spacecraft. Then, in order to realize realtime fault diagnosis of current spacecraft in orbit, parameters of the fault diagnosis model are readjusted to adapt the model to the current spacecraft fault diagnosis. Simulation results show that the proposed deep transfer learningbased fault diagnosis method can diagnose spacecraft fault signal quickly and accurately.

Key words: deep transfer learning, fault diagnosis, joint distribution adaption, residual network