Machine Learning-enhanced Reliability Analysis of a Cable-stayed Bridge subject to Stochastic Excitation

Abstract

As an important role in transportation networks, the slender cable-stayed bridges are inevitable to be subjected to stochastic excitation, which directly affects structural safety. Therefore, structural reliability analysis is great importance of structure design and safety evaluation in civil engineering. However, due to the structural complexity and highly nonlinear performance function, the traditional reliability analysis for cablestayed bridges is complicated and computationally intensive. To solve this problem, a machine learning-enhanced method of combining the back propagation (BP) neural network and subset simulation (SS) was developed in this report. Firstly, a 10-story shear building under seismic excitation was used to verify the proposed approach, which shows that the method has higher precision and can be applied to actual structure. Then, this thesis takes a cable-stayed bridge as the research object to conduct the wind-induced vibration reliability analysis. Consequently, the structural finite element model was established to calculate the wind-induced responses. To avoid complex calculation and reduce computational burden, the BP neural network was built as a surrogate model to predict the dynamic responses. Finally, the structural failure probability was estimated based on Monte Carlo simulation (MCS) and SS method, considering the uncertainty of elastic modulus, mass of main girder and wind speed. The results show that the BP neural network can achieve good prediction accuracy with the error of 0.29% to improve efficiency, and the machine learning-enhanced SS method has higher accuracy than MCS for the same number of samples.