No abstract

No abstract

Surrogate models are used to alleviate the computational burden in engineering tasks, which require the repeated evaluation of computationally demanding models of physical systems, such as the efficient propagation of uncertainties.For models that show a strongly non-linear dependence on their input parameters, standard surrogate techniques, such as polynomial chaos expansion, are not sufficient to obtain an accurate representation of the original model response.It has been shown that for models with discontinuities or rational dependencies, for example, frequency response functions of dynamic systems, the use of a rational (Padé) approximation can significantly improve the approximation accuracy. In order to avoid overfitting issues in previously proposed standard least squares approaches, we introduce a sparse Bayesian learning approach to estimate the coefficients of the rational approximation. Therein the linearity in the numerator polynomial coefficients is exploited and the denominator polynomial coefficients as well as the problem hyperparameters are determined through type-II-maximum likelihood estimation. We apply a quasi-Newton gradient-descent algorithm to find the optimal denominator coefficients and derive the required gradients through application of CR-calculus. The method is applied to the frequency response functions of an algebraic frame structure model as well as that of an orthotropic plate finite element model.

Several methods to localise sources of vibrations have been established in the literature. A great amount of those methods are based on databases with features of known impact positions. Great effort needs to be put into highly expensive experiments that deliver those databases. In this paper, we propose several simulation techniques that may replace the expensive experiments for source localisation. The paper compares the localisation accuracy of simulated and experimental data for two different localisation approaches, the reference database method and neural networks. Both methods process signal arrival time differences from several positions on the structure. The methods are exemplarily applied to a complex small-scale structure from the automotive industry: The small dimensions of the brake disk hat and the inclusion of holes is a challenging task for the accuracy of the applied localisation techniques. Results show that simulated data can replace experimentally gained data well in case of the reference database method, whereas the neuronal networks approach should stick to experimentally gained data. The evaluations show that, despite the small dimension, the relative localisation accuracy is within accepted ranges of literature.