Deep Transfer Learning-based Fault Diagnosis of Spacecraft Attitude System

In recent years, an increasingly popularity of deep learning model for intelligent state monitoring, diagnosis and prediction of spacecraft has been observed. However, in the previous studies, a major assumption accepted by default is that source domain data and target domain data have same feature distribution. Unfortunately, this assumption is mostly invalid in real application. Considering the problem that the original fault data sample is small, the noise is high and the fault signal is unlabeled, in this paper, we propose deep transfer learning-based fault diagnosis method for spacecraft system in which a new fault diagnosis framework-deep transfer network(DTN) is built, and it can generalize deep learning model to domain adaptation scenario inspired by the idea of transfer learning. In order to improve the accuracy of on-orbit spacecraft fault data detection, the proposed framework with joint distribution adaptation(JDA) is applied to exploit the distribution structure of unlabeled data in target domain by using the data with labels in source domain. By comparing with other methods, the deep transfer network based on joint distribution adaptation has better transfer performance in fault diagnosis of spacecraft.