Predicting bridge longitudinal displacement from monitored operational loads with hierarchical CNN for condition assessment

Over the past few years, longitudinal displacement has gained popularity as a means of evaluating the condition of long-span cable-supported bridge components, such as bearings. However, accurately predicting bearing displacement under varying load conditions is challenging due to the exposure of bridges to environmental and traffic loads. To address this issue, a hierarchical convolutional neural network (HCNN) model was developed in this paper for predicting bearing displacement using comprehensive loads as predictors. Structural health monitoring (SHM) systems of a cable-stayed bridge are utilized to provide one-month datasets for training and testing of the proposed method. Temperature, wind, and vehicle loads are adopted as input variables, and bearing displacement is the output. Results demonstrated the effectiveness of the proposed approach in predicting bearing displacement with an accuracy of over 95.6%, surpassing other models like traditional CNN, encoder-decoder, and U-Net in both accuracy and efficiency. Additionally, the contributions of different loads in predicting displacement are investigated, demonstrating the importance of traffic loads. Cumulative displacement can consequently be calculated for condition assessment of components such as bearings and expansion joints. A comparison with another cable-stayed bridge showed that the expansion joints in the current bridge were in satisfactory condition. Overall, the proposed approach can facilitate predictive maintenance in long-span bridges, helping to prevent premature failures.