Purpose
– A decentralized dynamic neural network (DNN)-based fault detection (FD) system for the reaction wheels of satellites in a formation flying mission is proposed. The paper aims to discuss the above issue.
Design/methodology/approach
– The highly nonlinear dynamics of each spacecraft in the formation is modeled by using DNNs. The DNNs are trained based on the extended back-propagation algorithm by using the set of input/output data that are collected from the 3-axis of the attitude control subsystem of each satellite. The parameters of the DNNs are adjusted to meet certain performance requirements and minimize the output estimation error.
Findings
– The capability of the proposed methodology has been investigated under different faulty scenarios. The proposed approach is a decentralized FD strategy, implying that a fault occurrence in one of the spacecraft in the formation is detected by using both a local fault detector and fault detectors constructed specifically based on the neighboring spacecraft. It is shown that this method has the capability of detecting low severity actuator faults in the formation that could not have been detected by only a local fault detector.
Originality/value
– The nonlinear dynamics of the formation flying of spacecraft are represented by multilayer DNNs, in which conventional static neurons are replaced by dynamic neurons. In our proposed methodology, a DNN is utilized in each axis of every satellite that is trained based on the absolute attitude measurements in the formation that may nevertheless be incapable of detecting low severity faults. The DNNs that are utilized for the formation level are trained based on the relative attitude measurements of a spacecraft and its neighboring spacecraft that are then shown to be capable of detecting even low severity faults, thereby demonstrating the advantages and benefits of our proposed solution.
Fault detection and isolation of satellite gyroscopes using relative positions in formation flying
