Using measurements to reduce model uncertainty for better predictions

Accurate models of real behaviour that are determined through measurements help engineers avoid expensive interventions and structural replacement. Model calibration by “curve-fitting” measurements to predictions is not appropriate for full-scale structures. This paper compares two population methods that can include modelling and measurement uncertainties using a simple example of a one-span beam. Standard applications of Bayesian inference that involve assumptions of independent zero-mean Gaussian distributions may not lead to accurate predictions, particularly when extrapolating. Another method, error-domain model falsification provides more reliable, albeit more approximate, predictions – especially when prediction is extrapolation. An example of a full-scale bridge illustrates the usefulness of the methodology in a real situation through improvements to fatigue-life estimates compared with design-type calculations without measurements.

Robust system identification and model predictions in the presence of systematic uncertainty

Romain Pasquier, Ian Smith

Model-based data-interpretation techniques are increasingly used to improve the knowledge of complex system behavior. Physics-based models that are identified using measurement data are generally used for extrapolation to predict system behavior under other actions. In order to obtain accurate and reliable extrapolations, model-parameter identification needs to be robust in terms of variations of systematic modeling uncertainty introduced when modeling complex systems. Approaches such as Bayesian inference are widely used for system identification. More recently, error-domain model falsification (EDMF) has been shown to be useful for situations where little information is available to define the probability density function (PDF) of modeling errors. Model falsification is a discrete population methodology that is particularly suited to knowledge intensive tasks in open worlds, where uncertainty cannot be precisely defined. This paper compares conventional uses of approaches such as Bayesian inference and EDMF in terms of parameter-identification robustness and extrapolation accuracy. Using Bayesian inference, three scenarios of conventional assumptions related to inclusion of modeling errors are evaluated for several model classes of a simple beam. These scenarios are compared with results obtained using EDMF. Bayesian model class selection is used to study the benefit of posterior model averaging on the accuracy of extrapolations. Finally, ease of representation and modification of knowledge is illustrated using an example of a full-scale bridge. This study shows that EDMF leads to robust identification and more accurate predictions than conventional applications of Bayesian inference in the presence of systematic uncertainty. These results are illustrated with a full-scale bridge. This example shows that the engineering knowledge necessary to perform parameter identification and remaining-fatigue-life predictions of a complex civil structure is easily represented by the EDMF methodology. Model classes describing complex systems should include two components: (1) unknown physical parameters that are identified using measurements; (2) conservative modeling error estimations that cannot be represented only as uncertainties related to physical parameters. In order to obtain accurate predictions, both components need to be included in the model-class definition. This study indicates that Bayesian model class selection may lead to over-confidence in certain model classes, resulting in biased extrapolation.

Elsevier2015

Robust system identification and model predictions in the presence of systematic uncertainty

Romain Pasquier, Ian Smith

Elsevier2015