Integration of graphical, physics-based, and machine learning methods for assessment of impact and recovery of the built environment from wind hazards
Date
2019
Authors
Pilkington, Stephanie F., author
Mahmoud, Hussam, advisor
Ellingwood, Bruce, committee member
van de Lindt, John, committee member
Zahran, Sammy, committee member
McAllister, Therese, committee member
Hamideh, Sara, committee member
Journal Title
Journal ISSN
Volume Title
Abstract
The interaction between a natural hazard and a community has the potential to result in a natural disaster with substantial socio-economic losses. In order to minimize disaster impacts, researchers have been improving building codes and exploring further concepts of community resilience. Community resilience refers to a community's ability to absorb a hazard (minimize impacts) and "bounce back" afterwards (quick recovery time). Therefore, the two main components in modeling resilience are: the initial impact and subsequent recovery time. With respect to a community's building stock, this entails the building damage state sustained and how long it takes to repair and reoccupy that building. In modeling these concepts, probabilistic and physics-based methods have been the traditional approach. With advancements in artificial intelligence and machine learning, as well as data availability, it may be possible to model impact and recovery differently. Most current methods are highly constrained by their topic area, for example a damage state focuses on structural loading and resistance, while social vulnerability independently focus on certain social demographics. These models currently perform independently and are then aggregated together, but with the complex connectivity available through machine learning, structural and social characteristics may be combined simultaneously in one network model. The popularity of machine learning predictive modeling across multiple different applications has risen due to the benefit of modeling complex networks and perhaps identifying critical variables that were previously unknown, or the mechanism behind how these variables interacted within the predictive problem being modeled. The research presented herein outlines a method of using artificial neural networks to model building damage and recovery times. The incorporation of graph theory to analyze the resulting models also provides insight into the "black box" of artificial intelligence and the interaction of socio-technical parameters within the concept of community resilience. The subsequent neural network models are then verified through hindcasting the 2011 Joplin tornado for individual building damage and the time it took to repair and reoccupy each building. The results of this research show viability for using these methods to model damage, but more research work may be needed to model recovery at the same level of accuracy as damage. It is therefore recommended that artificial neural networks be primarily used for problems where the variables are well known but their interactions are not as easily understood or modeled. The graphical analysis also reveals an importance of social parameters across all points in the resilience process, while the structural components remain mostly important in determining the initial impact. Final importance factors are determined for each of the variables evaluated herein. It is suggested moving forward, that modeling approaches consider integrating how a community interacts with its infrastructure, since the human components are what make a natural hazard a disaster, and tracing artificial neural network connections may provide a starting point for such integration into current traditional modeling approaches.
Description
Rights Access
Subject
building recovery
graph theory
wind damage
community resilience
artificial neural networks
socio-technical