Development of a high-fidelity failure prediction system for reinforced concrete bridge columns using generative adversarial networks

Abstract

Visual damage patterns of structural components are critical to evaluate structural performance. However, visual damage images can only be obtained by laboratory tests or in-situ tests, which require massive funding and human workforces. In this study, a novel bridge column failure prediction framework is proposed based on generative adversarial network (GAN). Trained with merely 110 damage patterns collected from experimental cyclic loading test, the proposed approach predicts the high-fidelity surface damage patterns of concrete bridge columns given the information of the column design parameters as well as the user-desirable performance level, i.e., Damage Index (DI) of the column. Two network architectures and three label encoding strategies are explored to investigate the performance in estimating the damage pattern. By incorporating DI as a numerical label, the proposed network is able to predict the unseen damage patterns which are not available in the training dataset. Also, it is found that adding classifiers and regressors in the discriminator to account for the condition vector is beneficial for network training, achieving a Frechet Inception Distance (FID) of 102.6 when producing the synthetic patterns. Extensive experiments have demonstrated that the proposed framework is capable of synthesizing decent damage patterns with superior fidelity, providing bridge engineers with a platform to evaluate the potential failure modes during seismic design and evaluation.