Buildings - Special Topics in Structures

Full Session with Abstracts

339370-3 - Supervised Machine Learning Approaches for Efficient Reliability Analysis of Structural and Infrastructure Systems

Saturday, April 21
8:00 AM - 9:30 AM
Location: 202A

Reliability analysis is a key step in probabilistic evaluation of the hazard performance of structural and infrastructure systems. The assessment of reliability in these areas often involves simulations of system responses using sophisticated numerical models. Considering that failure probabilities of these systems are very small, conventional reliability analysis techniques such as Monte Carlo simulation (MCS) and first and second order reliability methods (FORM and SORM) either require a significantly large number of simulations to converge or they fail to achieve acceptable levels of accuracy. Hence, it is critical to develop strategies that can achieve high accuracy in estimation of failure probabilities of reliable systems with reduced number of calls to performance models.
Machine learning algorithms such as supervised learning methods can assist in achieving this goal by establishing a surrogate model that is derived through treatment of realizations of random variables and corresponding simulation-based responses as observations. Developed based on a limited number of observations, these surrogate models can then substitute the original complex and time-consuming performance models in the reliability estimation process. After these models are established, they can be integrated with simulation-based reliability analysis methods such as MCS and importance sampling techniques. The application of a state-of-the-art supervised learning based algorithm called adaptive Kriging–based MCS (AK-MCS) in estimating the reliability of structural systems is explored here. This paper also extends AK-MCS by incorporating error rate estimates in the selection of design points; this new algorithm is called Reliability analysis with Error rate-based Adaptive Kriging (REAK). Two examples including a complex analytical limit state function and a truss bridge are considered here. The reliability of these structures is estimated using conventional techniques of MCS and FORM, as well as AK-MCS and REAK. These methods are compared with respect to the accuracy of their reliability estimates, the number of calls to performance models, and the rate of convergence. Results have indicated that reliability estimates by FORM may not have sufficient accuracy for nonlinear and high-dimensional limit state functions. Moreover considering that a call to a complex performance function may require several hours of simulation time, MCS is not a practical solution for reliability analysis of complex systems. On the other hand, it is observed that the machine learning based AK-MCS and the proposed methodology REAK provide high accuracy in reliability estimates with a limited number of calls to performance models. The proposed algorithm appears to be the most efficient method with the least number of calls to performance models.
The presented algorithms based on machine learning techniques provide practical and reliable solutions for estimation of failure probabilities of structural and infrastructure systems. These methods will help researchers with hazard performance assessment of various complex systems as they significantly reduce the number of calls to time-consuming performance models. Moreover, the high accuracy of reliability estimates by these methods enhances risk assessment of infrastructure systems and leads to more optimal solutions for their risk management.

Abdollah Shafieezadeh

Assistant Professor
The Ohio State University

Critical infrastructure systems (CIs) are becoming increasingly vulnerable to system-wide failure primarily due to aging, natural and manmade hazards, climate change, and increased interdependencies, among other factors. Quantitative and qualitative assessment of the associated risk is crucial in pre-event planning and post-event response. Risk assessment and management of structural and infrastructure systems (RAMSIS) lab develops and applies probabilistic risk analysis frameworks to various CIs and their components to assess their reliability and resilience against perturbations. This can provide valuable knowledge about a number of principal effects such as traffic disruption, impact on the regional and global economy, and resilience of systems and communities. The focus of Dr. Shafieezadeh's research is on numerical modeling and fragility assessment of complex systems, such as seaports, bridges, and physical assets of power grids with consideration of soil-foundation-structure interactions and liquefaction effects, probabilistic modeling of deterioration processes of reinforced and prestressed concrete structures coupled with FE simulations, optimal maintenance policies for large aging infrastructure assets using stochastic methods, advanced protection of critical geo-structural systems against extreme hazards such as earthquakes and strong winds using passive, active, and semi-active control strategies based on stochastic methods, reliability and resilience assessment of geo-structural systems and critical infrastructures against natural and manmade hazards.

Presentation(s):

Send Email for Abdollah Shafieezadeh


Assets

339370-3 - Supervised Machine Learning Approaches for Efficient Reliability Analysis of Structural and Infrastructure Systems



Attendees who have favorited this

Please enter your access key

The asset you are trying to access is locked. Please enter your access key to unlock.

Send Email for Supervised Machine Learning Approaches for Efficient Reliability Analysis of Structural and Infrastructure Systems