Events Calendar

Astani Civil and Environmental Engineering Ph.D. Seminar

Speaker: Subhayan De and Simin Karvigh , Astani CEE Ph.D. Candidates

Talk Title: Efficient Bayesian Model Selection for Locally Nonlinear Systems incorporating Dynamic Measurements

Abstract: Subhayan De's abstract:

The modeling of a structural system is often complicated by the dynamic characterization of a component by competing families of models, also known as model classes, where the choice of a particular model class falls to the discretion of the researcher. Bayesian model selection can be used to help find the most plausible model class. For linear models, the computational effort for characterization of dynamic properties using natural frequencies and mode shapes, as well as Monte Carlo sampling method, is reasonably well understood. On the other hand, to characterize the dynamic behavior of nonlinear models, response time histories are needed, resulting in high computational cost even when most of the structure is linear and the nonlinear behavior is very localized.

In this study, the computational effort of Bayesian model selection is dramatically reduced in two ways: (1) using a more intelligent Monte Carlo sampling and (2) exploiting the local nature of the nonlinearities. The marginal likelihoods, which are the evidences of the model classes, are estimated with response time histories using ‘nested sampling’ (Skilling 2006), which samples more from regions with high likelihood values than regions with low likelihood regions. The localized nature of the nonlinearities in the dynamic system is exploited using an efficient response calculation algorithm (Gaurav et al. 2011) by transforming the system equations of motion to a low-order nonlinear Volterra integral equation (NVIE) that is solved numerically. This approach is demonstrated with numerical models of a base-isolated 11-story 2-bay 99-DOF superstructure on the hysteretic lead rubber bearing (LRB) isolators. Model selection is performed to choose from among six model classes: four linear (AASHTO, CALTRANS, JPWRI, modified AASHTO) and two nonlinear (Bouc-Wen, bilinear) models of the isolator, using simulated responses to historical earthquake records. The computational efficiency of the proposed approach is compared with a traditional ordinary differential equation solver (MATLAB’s ode45) demonstrating speedup up to two orders of magnitude.

Location: Seeley G. Mudd Building (SGM) - 101

Audiences: Everyone Is Invited

Contact: Evangeline Reyes