Place: GKSS Forschungszentrum Geesthacht, Germany
Norbert Huber, Technische Universität Hamburg-Harburg, Hamburg, Germany
Articial neural networks can be used to solve complex inverse problems, e.g. the identication so the stress-strain behaviour from spherical indentation. The talk will discuss a possible application of this method for determining local mechanical properties in welded components, as they are relevant for light-weight design in an aerospace fuselage. Welded structures always have a residual biaxial stress state in the welded area with a major component in longitudinal direction. For more than a decade we know that residual stresses can have a strong eect on the result of an indentation experiment. The eect of residual stresses can generally be simu- lated in experiments by simple bending (uniaxial stress state) and in nite element simulations by application of transverse stresses as a boundary condition (typically equibiaxial). Many researchers have used these approaches to develop models that allow the prediction of this eect; others have proposed models to determine resid- ual stresses from indentation or hardness testing. The training of articial neural networks (ANNs) allows also analysing the relationship between independent (stress state) and dependent quantities (load, hardness) using a large number of nite ele- ment simulations. Based on the experience gathered from the ANN results and nite element studies, a simple analytical model was developed, which allows predicting the eect of an arbitrary residual stress state on the measured hardness with an accuracy of better than 5%. Using this model it is easily possible to estimate errors in hardness measurement of to correct hardness data if the in-plane residual stresses can be measured.