Neural Network-based Actuator Fault Diagnosis for Attitude Control Subsystem of an Unmanned Space Vehicle

The main objective of this paper is to develop a neural network-based fault detection and isolation scheme (FDI) for the attitude control subsystem (ACS) of a satellite. Towards this end, two neural network architectures are considered. First, a dynamic neural network residual generator is constructed based on the dynamic multilayer perceptron (DMLP) network to perform the detection task. A generalized embedded structure for the dynamic neuron model is considered in the DMLP network. Second, a static neural classifier is developed based on learning vector quantization (LVQ) network to be utilized for the isolation task. Based on a given set of input-output data collected from a 3-axis ACS of a satellite, the network parameters are adjusted to minimize a performance index specified by the output estimation error. The proposed neural FDI structure is applied to detect and isolate various faults in a high-fidelity nonlinear model of a satellite reaction wheel (RW), which is often used as an actuator in the ACS. The performance and capabilities of the proposed techniques are investigated and compared to a model-based observer residual generator that is used to detect various fault scenarios.