Integrated Framework for Assessment of Time-Variant Flood Fragility of Bridges Using Deep Learning Neural Networks

Abstract

The increasing intensity and frequency of extreme hydrological events poses a significant threat to the safety of transportation infrastructure across the globe. To reduce the failure risk associated with these structures, bridge assessment and management approaches should adapt to possible increases in future flood hazards. Fragility analysis can assist infrastructure managers in properly quantifying the reliability of bridges under different flood hazard intensity levels. However, conducting such analysis while accounting for various uncertainties associated with bridge capacity, deterioration, and future climate conditions can significantly increase the computational cost of bridge management procedures. To improve the computational efficiency of the fragility analysis while maintaining the desired accuracy, this paper integrates deep learning (DL) neural networks in a simulation-based probabilistic framework for establishing time-variant fragility surfaces of bridges under flood hazard. The proposed probabilistic framework considers the effects of climate change on flood occurrence and long-term scour hazard. Downscaled climate data, adopted from global climate models, are used to predict future precipitation and temperature profiles at a given location. Deep learning networks are employed with a twofold objective: (1) to predict future river streamflow at an investigated location necessary for assessing the scour conditions and flood hazard at the bridge, and (2) to simulate the structural behavior of a bridge foundation under sour conditions. The trained DL networks are integrated into a probabilistic simulation process to quantify failure probability and construct a bridge fragility surface under flood hazard. The proposed framework is illustrated on an existing bridge located in Oklahoma.