Zuverlässigkeitsanalyse und -Aktualisierung komplexer Infrastruktursysteme mit Bayes'schen Netzwerken
Zusammenfassung der Projektergebnisse
Civil infrastructure systems such as transportation or energy networks are critical backbones of modern societies and economies. To understand and enhance their hazard resilience, it is essential to quantify the system-level risk and reliability by use of models and data from multiple fields of science and engineering. Ideally, such risk estimates are updated when new information becomes available during the lifetime of the system. As an example, the risk estimates should be updated with available data immediately after a disaster event such that decisions on hazard response and post-disaster network operations can be made promptly but prudently. The Bayesian network (BN) methodology has been identified as a powerful tool that can facilitate the integration of models and data from different areas and has previously been proposed for near-real-time reliability assessment of infrastructure systems during and after hazard events. In this project we worked on enhancing models and algorithms to overcome computational limitations of existing BN algorithms in predicting the reliability of infrastructure systems. Different approaches were investigated, resulting in two key outcomes: The first is a modeling and computation framework that combines the matrix-based system reliability (MSR) method with BN, termed Matrix-based Bayesian network (MBN). The second is an algorithm for efficient samplingbased BN evaluation targeted towards reliability problems, which combines BN with subset simulation (SuS). The MBN introduces an efficient representation of the infrastructure system that can be used in a BN structure. The MBN uses conditional probability matrices (CPMs) as an alternative data structure, which consist of a set of rules that assign a system state to a set of input states and an associated probability. This data structure is significantly more effective than the classical conditional probability tables in BN for modeling infrastructure systems. It has the additional advantage that it is also possible to specify the rules only incompletely, in which case the MBN computes lower and upper bounds of the system reliability. Together with the data structure, we developed associated algorithms that enable the use of the standard BN algorithms with the MBN data structure. The developed BN-SuS algorithm is based on a SuS algorithm for the system performance node in combination with an MCMC sampler for the remaining BN. In particular, a Gibbs sampler is used, within which a combination of a slice sampler and a univariate Metropolis-Hastings (MH) sampler is employed. These samplers were optimized to facilitate the evaluation of the intermediate conditional probabilities that are the basis of SuS. The proposed algorithm enables efficient computation of large systems with many components under evidence from different sources. All algorithms were tested and verified on a number of example applications considering random networks as well as transportation and gas networks.
Projektbezogene Publikationen (Auswahl)
-
(2017). Enhancing sampling based inference in hybrid BNs for reliability assessment. Proc. ICOSSAR 2017, Vienna
Zwirglmaier K., Papaioannou I., Straub D.
-
(2018). An Improved Non-parametric Bayesian Independence Test for Probabilistic Learning of the Dependence Structure among Continuous Random Variables. KSCE Journal of Civil Engineering, 22(3), 974-986
Byun J., Song J., Zwirglmeier K., Straub D.
-
(2019). Hybrid Bayesian networks for modeling large infrastructure systems. Proc. 29th European Safety and Reliability Conference ESREL, Hanover
Zwirglmaier K., Straub D.
-
(2019). Matrix-based Bayesian Network for efficient memory storage and flexible inference. Reliability Engineering & System Safety, 185: 533-545
Byun J.-E., Zwirglmaier K., Straub D., Song J.
