Fault Location for spacecraft ACS system using the method of transfer learning

As one of the important subsystems of the spacecraft, the attitude control system helps to maintain the stability of spacecraft and carry out space tasks. Thus, it is very meaningful to accomplish its fault location methods to maintain the working order of the spacecraft. The data-driven method is the mainstream method of fault diagnosis at present, because there are a lot of telemetry data, which can reflect the in-orbit operation information of the spacecraft. The current data-driven methods always need fault samples. However, fault samples are often difficult to obtain, which brings difficulties and challenges to the application of data-driven methods. This paper proposes an improved fault location method by introducing transfer learning to it, with which the data structure of source domain learned by BP neural network could be transferred to the target domain. Through theoretical analysis, When the condition is satisfied that difference between source domain model and target domain model is within a certain range, this method could realize the fault location of spacecraft system without fault samples. Finally, a semi-physical simulation test was carried out based on a micro triaxial air bearing table simulation system in order to verify the effectiveness of the proposed method. The results show that under the condition of lack of fault samples, the proposed method can significantly improve the accuracy of ACS fault location compared with the traditional BP neural network method.