Investigating the Generalization Capability and Performance of Neural Networks and Neuro-Fuzzy Systems for Nonlinear Dynamics Modeling of Impaired Aircraft

Abstract

One of the challenging problems in the case of aircraft failure is to determine the new altered dynamics of the impaired aircraft. Among various methods, neural networks and neuro-fuzzy systems can be used for high-fidelity modeling of the aircraft nonlinear dynamics with the aim of onboard applications in real time. However, the method with better generalization capability is more preferred specifically in the case of unpredicted aircraft failures. Generalization of a network is mainly dependent on the network's parameters, the employed training algorithm, and the amount of training data. In this paper, several neural networks and local model networks are trained using different training algorithms and different amounts of training data to model the nonlinear dynamics of an impaired aircraft with the damaged rudder. These networks are compared based on their generalizations to the new cases of rudder failure. The effect of using different amounts of training data on the generalization capability and performance of the networks has also been investigated. The results of this paper show that both network types have good performance but neural networks generalize better to the new failure cases than local model networks. Also based on the obtained results, a significant reduction in the number of training samples could be accomplished without a considerable decrease in the network's performance and generalization. Finally, a neural network-based sensitivity analysis method is proposed which utilizes the network's regression equation as an emulator for fast model evaluations and can be used as an advisory tool for choosing safer path planning strategies.