Abstract: The data-based fault diagnosis method has become a key research direction in the current fault diagnosis field due to its excellent engineering applicability; but the training of its algorithm model generally requires sufficient sample data, so it is difficult to solve the diagnosis problem lacking of fault samples. Aiming at the problem of satellite attitude control system (ACS) fault diagnosis when the target satellite has no fault samples, a fault location technology based on DDC (deep domain confusion) transfer learning algorithm improvement is proposed. Firstly, nominal satellite attitude information is reconstructed by the longshort term memory auto encoder (LSTM-AE) network and its residuals are calculated. Secondly, the features of extraction are trained into the BP network fault location classifier. Thirdly, the DDC transfer learning algorithm is introduced: a domain adaptation layer and modification of the loss function are added to classifier network to learn the health and fault characteristics of the target satellite and improve the algorithm model. Finally, through the semi-physical simulation platform of the threeaxis air bearing table, the effectiveness of the fault location technology improved by the introduction of DDC migration learning is verified.

Key words: transfer learning, neural networks, LSTM-AE, fault location