Probabilistic prediction on long-term structural responses of cable bridges using Bayesian neural networks

Abstract

Various studies have introduced methods for identifying the structural conditions of cable-stayed or suspension bridges using prediction models. It is essential to improve both the accuracy and efficiency of these models for structural assessment. To achieve this, existing research suggests that incorporating seasonal or monthly fluctuations into long-term response predictions can yield more accurate results by accounting for corresponding variations in the structural response. However, when predicting structural responses using machine learning models and measurement data, it is crucial to account for uncertainty. Even with larger training datasets, effectively reducing uncertainty remains a challenge. Therefore, this study proposes a probabilistic approach for predicting long-term structural responses, aiming to enhance both accuracy and efficiency. Additionally, the proposed method can identify the structural condition by detecting anomalies in measurement data from stayed cables on a cable-stayed bridge. Given the large number of long-term monitoring datasets to be considered for prediction, this study employs Bayesian Neural Networks (BNNs)-a deep learning model-to further improve efficiency. Finally, to verify the performance of the BNNs, the method is applied to artificially generated measurement data (i.e., synthetic data). Through this application example, BNNs are demonstrated to be accurate and efficient in identifying cable conditions.