Combined Geometric and Neural Network Approach to Generic Fault Diagnosis in Satellite Reaction Wheels

This paper suggests a novel diagnosis scheme for detection, isolation and estimation of faults affecting satellite reaction wheels. Both spin rate measurements and actuation torque defects are dealt with. The proposed system consists of a fault detection and isolation module composed by a bank of residual filters organized in a generalized scheme, followed by a fault estimation module consisting of a bank of adaptive estimation filters. The residuals are decoupled from aerodynamic disturbances thanks to the Nonlinear Geometric Approach. The use of Radial Basis Function Neural Networks is shown to allow design of generalized fault estimation filters, which do not need a priori information about the faults internal model. Simulation results with a detailed nonlinear spacecraft model, which includes disturbances, show that the proposed diagnosis scheme can deal with faults affecting both reaction wheel torques and flywheel spin rate measurements, and obtain precise fault isolation as well as accurate fault estimates.