Local Sensitivity Analysis of Nonlinear Models - Applied to Aircraft Vehicle Systems

Abstract: As modeling and simulation becomes a more important part of the modeling process, the demand on a known accuracy of the results of a simulation has grown more important. Sensitivity analysis (SA) is the study of how the variation in the output of a model can be apportioned to different sources of variation. By performing SA on a system, it can be determined which input/inputs influence a certain output the most. The sensitivity measures examined in this thesis are the Effective Influence Matrix, EIM, and the Main Sensitivity Index, MSI. To examine the sensitivity measures, two tests have been made. One on a laboratory equipment including a hydraulic servo, and one on the conceptual landing gear model of the Gripen aircraft. The purpose of the landing gear experiment is to examine the influence of different frictions on the unfolding of the landing gear during emergency unfolding. It is also a way to test the sensitivity analysis method on an industrial example and to evaluate the EIM and MSI methods. The EIM and MSI have the advantage that no test data is necessary, which means the robustness of a model can be examined early in the modeling process. They are also implementable in the different stages of the modeling and simulation process. With the SA methods in this thesis, documentation can be produced at all stages of the modeling process. To be able to draw correct conclusions, it is essential that the information that is entered into the analysis at the beginning is well chosen, so some knowledge is required of the model developer in order to be able to define reasonable values to use. Wishes from the model developers/users include: the method and model quality measure should be easy to understand, easy to use and the results should be easy to understand. The time spent on executing the analysis has also to be well spent, both in the time preparing the analysis and in analyzing the results. The sensitivity analysis examined in this thesis display a good compromise between usefulness and computational cost. It does not demand knowledge in programming, nor does it demand any deeper understanding of statistics, making it available to both the model creators, model users and simulation result users.