Spacecraft Fault Location Method Based on DSAN

Abstract

AbstractAs space science and technology improves by leaps and bounds, the reliability evaluation of spacecraft has become a research focus in the space field. Therefore, the requirements of spacecraft fault diagnosis are increasing. Attitude control system (ACS) is the most critical part of satellite, and it is also the place where faults occur frequently. However, the fault data of spacecraft in orbit are insufficient and mostly unlabeled, which poses a great challenge to fault location. In this paper, a fault location method based on deep transfer learning combined with simulation system is proposed to solve the fault location problem of satellite attitude control system. Firstly, a simulation model was built by using PD control method to simulate the satellite ACS. Meanwhile, artificial fault injection was used to obtain health and fault data as the source domain. Furthermore, a large number of health data and a small amount of fault data of the air bearing table are collected as the target domain, and then the neural network is trained to simulate ACS for the health data of the source domain and target domain, so that the ACS output results are converted to grayscale images after residual difference with the operating data. Then, after comparing some transfer learning methods in Alexnet network, the Deep Subdomain Adaptive Network (DSAN) method is used to realize transfer, which enables the model to locate faults in the target domain with only a few fault samples (labeled and unlabeled) in the target domain. The results of simulation show that the model can quickly locate the fault.KeywordsAttitude control systemFault locationDeep transfer learningDeep subdomain adaptive network