Augmenting Bayesian Inference-Based Damage Diagnostics of Miter Gates Based on Image Translation

Abstract Structural monitoring data is essential for the assessment of structural integrity of large civil infrastructures, such as miter gates studied in this paper. For some applications, the amount of monitoring data is limited due to various reasons including but not limited to the availability of sensors, changing states of structural health, inaccessibility of measurement location, etc.. This limited amount of monitoring data could result in uncertainty in structural health assessment. This paper overcomes this challenging issue by proposing a data augmentation method based on image translation for Bayesian inference-based damage diagnostics of miter gates. In particular, we translate the monitoring data of one miter gate to that of another, thereby increasing the volume of monitoring data available for assessing the structural health of a target miter gate. This translation starts with converting the monitoring data of different miter gates into images. After that, Cycle Generative Adversarial Networks (CycleGAN) are employed to accomplish the task of image translation among different miter gates. The translated images (i.e., synthetic monitoring data) are then used together with the true monitoring data for damage diagnostics. Two types of CycleGAN architectures are investigated and compared using a case study. Results of the case study show that the proposed data augmentation method can effectively improve the accuracy and confidence of damage diagnostics of miter gates. It demonstrates the potential of integrating synthetic data generation with probabilistic model updating in structural health monitoring.