Filtern
Dokumenttyp
- Vortrag (2)
- Posterpräsentation (1)
Sprache
- Englisch (3)
Referierte Publikation
- nein (3)
Schlagworte
- Variational Bayesian (3) (entfernen)
Organisationseinheit der BAM
- 7 Bauwerkssicherheit (3) (entfernen)
Eingeladener Vortrag
- nein (2)
Appropriate monitoring of transportation infrastructures (e.g. bridges) is of utmost importance to ensure safe operation conditions. Accurate and reliable assessment of such structures can be achieved through the integration of data from non-destructive testing, advanced modeling and model updating techniques. The Bayesian framework has been widely used for updating engineering and mechanical models, due to its probabilistic description of information, in which the posterior probability distribution reflects the knowledge, over the model parameters of interest, inferred from the data. For most real-life applications, the computation of the true posterior involves integrals that are analytically intractable, therefore the implementation of Bayesian inference requires in practice some approximation methods.
This paper investigates the application of Variational Bayesian Inference for structural model parameter identification and update, based on measurements from a real experimental setup. The Variational Bayesian method circumvents the issue of evaluating intractable integrals by using a factorized approximation of the true posterior (mean field approximation) and by choosing a family of conjugate distributions that facilitates the calculations. Inference in the Variational Bayesian framework is seen as solving an optimization problem with the aim of finding the parameters of the factorized posterior which would minimize its Kullback-Leibler divergence in relation to the exact posterior. The Variational Approach is an efficient alternative to sampling methods, such as Markov Chain Monte Carlo, since the latter’s accuracy depends on sampling from the posterior distribution a sufficient amount of times (and therefore requiring an equivalent number of computations of the forward problem, which can be quite expensive).
Appropriate monitoring of transportation infrastructures (e.g. bridges) is of utmost importance to ensure safe operation conditions. Accurate and reliable assessment of such structures can be achieved through the integration of data from non-destructive testing, advanced modeling and model updating techniques. The Bayesian framework has been widely used for updating engineering and mechanical models, due to its probabilistic description of information, in which the posterior probability distribution reflects the knowledge, over the model parameters of interest, inferred from the data. For most real-life applications, the computation of the true posterior involves integrals that are analytically intractable, therefore the implementation of Bayesian inference requires in practice some approximation methods.
This paper investigates the application of Variational Bayesian Inference for structural model parameter identification and update, based on measurements from a real experimental setup. The Variational Bayesian method circumvents the issue of evaluating intractable integrals by using a factorized approximation of the true posterior (mean field approximation) and by choosing a family of conjugate distributions that facilitates the calculations. Inference in the Variational Bayesian framework is seen as solving an optimization problem with the aim of finding the parameters of the factorized posterior which would minimize its Kullback-Leibler divergence in relation to the exact posterior. The Variational Approach is an efficient alternative to sampling methods, such as Markov Chain Monte Carlo, since the latter’s accuracy depends on sampling from the posterior distribution a sufficient amount of times (and therefore requiring an equivalent number of computations of the forward problem, which can be quite expensive).
Numerical simulators, such as finite element models, have become increasingly capable of predicting the behaviour of structures and components owing to more sophisticated underlying mathematical models and advanced computing power. A common challenge lies, however, in calibrating these models in terms of their unknown/uncertain parameters. When measurements exist, this can be achieved by comparing the model response against measured data. Besides uncertain model parameters, phenomena like damage can give rise to further uncertainties; in particular, quasi-brittle materials, like concrete, experience damage in a heterogeneous manner due to various imperfections, e.g. in geometry and boundary conditions. This hardens an accurate prediction of the damaged behaviour of real structures that comprise such materials.
In this study, which draws from a data-driven approach, we use the force-version of the finite element model updating method (FEMU-F) to incorporate measured displacements into the identification of the damage parameters, in order to cope with heterogeneity. In this method, instead of conducting a forward evaluation of the model and comparing the model response (displacements) against the data, we impose displacements to the model and compare the resulting force residuals with measured reaction forces. To account for uncertainties in the measurement of displacements, we endow this approach with a penalty term, which reflects the discrepancy between measured and imposed displacements, where the latter is assumed as unknown random variables to be identified as well. A Variational Bayesian approach is used as an approximating tool for computing posterior parameters. The underlying damage model considered in this work is a gradient-enhanced damage model.
We first establish the identification procedure through two virtual examples, where synthetic data (displacements) are generated over a certain spatially-dense set of points over the domain. The procedure is then validated on an experimental case-study; namely a 3-point bending experiment with displacement measurements resulting from a digital image correlation (DIC) analysis.