Comparing Three Methodologies for System Identification and Prediction

Most civil infrastructure in service today was built during the second half of 20th century and is now reaching the end of its design life. Replacement of all aging civil infrastructure is a drain on national and global economies. Design models for civil infrastructure are justifiably conservative. Decision making related to asset management activities such as repair, improvement and extension of existing infrastructure can be enhanced through structural identification and capacity prediction. Recent advances in sensing and computing technologies enable use of model-based data interpretation methods, such as residual minimization, Bayesian model updating and error-domain model falsification (EDMF) for structural identification. In the traditional Bayesian-model-updating approach for parameter identification, the uncertainty is assumed to be defined by uncorrelated Gaussian distributions. However, in civil infrastructure, the uncertainty associated with the system is rarely Gaussian and often systematic with high, yet unknown, correlations. In this paper, a modified Bayesian model updating methodology with L_∞-norm-based likelihood function is compared with EDMF and traditional Bayesian methodology. The traditional Bayesian model updating methodology may provide biased prediction when correlations are unknown. The results obtained using the modified Bayesian model-updating approach are similar to the results obtained using EDMF. The three methodologies are compared with respect to their ease of integration of domain knowledge and their adaptability to changing information. Compared with traditional Bayesian model updating methodology, EDMF and modified Bayesian model updating methodologies provide robust, albeit less precise, prediction of structural response at unmeasured locations for civil-engineering infrastructure. Finally, EDMF has advantages over Bayesian methodologies for practical engineering use.