A Comparison Between Generalization Capability of Neural Network and Neuro-Fuzzy System in Nonlinear Dynamics Identification of Impaired Aircraft

In case of technical failures or external events such as control surface defects or icing, aircraft dynamics and parameters are changed. Due to nonlinear dynamics of aircraft, usually the exact new altered dynamics cannot be determined by the pilot. Therefore the pilot who tries to plan a safe landing trajectory as soon as possible may implement a maneuver which is not feasible anymore according to the altered dynamics of the impaired aircraft and leads to aircraft loss of control (LOC). Therefore, the main challenge in the prevention of LOC-led-accidents is to increase the pilot's situational awareness and develop better control systems which both require post-failure dynamics identification and modeling. Both neural networks and neuro-fuzzy systems can be used for high-fidelity modeling of the aircraft nonlinear dynamics, however, the one with better generalization capability should be chosen. In this paper, several neural networks and local model networks are developed to model the nonlinear dynamics of an impaired aircraft with damaged rudder. These networks are trained using different training algorithms and their generalizations to the new cases of rudder failure are compared. Results show that both network types have good performance but neural networks generalize better to the new data than local model networks.