Identification of nonlinear aerodynamic derivatives using classical and extended local model networks

Abstract

Determining aerodynamic models for use in simulators requires the model to be valid over a wide range of flight conditions. Local model networks are suitable for this kind of task because they build a global model through a weighted superposition of local simple models. The location of the local models, i.e. the partitioning into submodels is determined automatically as part of the algorithm. Unlike neural networks that yield only black-box models, the structure and parameters of local model networks are interpretable and can quite easily be transformed into modeling functions or table models. Using flight test data, it is shown that local model networks are useful in the identification of models that have to cover a broad range of flight conditions. When identi-fying aerodynamic parameters from flight test data, often the task is to derive models for the different nonlinear derivatives directly from measurements of the overall coefficient. For this, two extensions of the classical local model networks are introduced and investigated. Out of the two approaches, the structured local networks yield very promising results.