A Dynamic Neural Network-Based Fault Detection of Reaction Wheels in the Attitude Control Subsystem of Formation Flying Satellites

This paper proposes a methodology for detecting faults in any of the multiple reaction wheels that are employed in a consensus-based virtual structure controlled formation flight of satellites. The highly nonlinear dynamics of the formation flight and reaction wheels are modeled by using dynamic multilayer perceptron neural networks. The proposed dynamic neural networks (DNNs) utilize the extended back propagation learning algorithm and are trained based on sets of input/output data collected from the relative attitude determination sensors of the 3-axis attitude control subsystems of the satellites. The DNN parameters are adjusted to minimize the performance indices (representing the output estimation error). The capability of the proposed DNN is investigated under various faulty situations, including single and multiple actuator fault scenarios.