Fast-GLP: a Fast Tool for the Prediction of Worst Case Gust Loads based on Neural Networks

Abstract

This paper discusses a procedure developed in the framework of FFAST (Future Fast Aeroelastic Simulation Technologies) project aiming to significantly reduce the amount of time required to determine the worst case gust loads for aircraft. The tool is based on the coupling between an efficient gust response solver with a Neural Network metamodel and aims at minimizing the number of analysis cases across the entire flight envelope. The results concerning the application of the proposed approach to a typical large transport aircraft are shown to demonstrate the accurate predictions of the worst gust case with a limited set of gust response calculations.

Parametric reduced order model approach for rapid dynamic loads prediction

“…Note, however, that the errors relating to higher frequencies are caused by the balanced truncation itself rather than by the interpolation of the locally reduced models. The approximation errors are in line with those presented in [7], [8], where Neural Networks and system identification were used as black-box surrogates to approximate the extreme responses. However, following those approaches, a different metamodel must be built for each IQ, a cost that could quickly become unacceptable for an industrial case where thousands of IQs are monitored.…”

“…A considerable saving in computational effort can be made if, for the thousands of simulations required during an aircraft loads loop or for quantification of the effects of parameter uncertainties on the aeroelastic behaviour, a ROM is used in place of the high dimensional model. The ROM could thus be seen as a physics-based surrogate alternative to the data-fitting approaches, such as Kriging, Radial Basis Functions, Neural Networks or system identification proposed for the same purpose in [6], [7], [8]. Whereas a data-fit surrogate model, created in a black-box mode, maps an input/output relationship, a ROM embodies the underlying physics of the problem and, unlike the aforementioned methods, its validity is not limited to the conditions under which it was generated, but can be applied to simulate various initial conditions.…”

“…The time domain approaches defined in the airworthiness regulations [1], [5] are based upon finding the tuned "1 -Cosine" gust that causes the worst response. Previous work in the FFAST EU FP7 project [6], [7], [8] investigated the use of several surrogate models and optimization methods for fast and efficient prediction of the worst case gust loads for each IQ of a large transport aircraft model. It was shown that considerable savings in computational time can be made without sacrificing accuracy; however, the IQs were dealt with independently.…”

“…Generalized aerodynamic forces matrix due to control surfaces Pilot command input 4 56 Weighting factor of the optimization process for the extended Rational Function Approximation 4 7 Gust Aerodynamic poles of Rational Function Approximation < Control surface deflection = 56 Normalized squared error between tabulated and approximated i-j terms of generalized aerodynamic matrix >?…”

Preliminary aeroelastic design of composite wings subjected to critical gust loads

“…Knoblach [8] used robust performance analysis from control theory to identify critical loads due to discrete 1-cosine gusts. In the work done under the European FP7 project FFAST [5,9,10], surrogate modelling, neural networks and optimization techniques were used for fast prediction of gust loads.…”