Examination of load-deformation characteristics of long-span bridges in harsh natural environments based on real-time updating artificial neural network

Long-span bridges, often exposed to challenging harsh natural environments with severe weather conditions, necessitate real-time examination of load-deformation characteristics to ensure structural integrity and safety. Previous studies have primarily focused on investigating the causes of deformation in bridge structures under different single-load conditions during severe natural disasters, utilizing physics-based, mechanics-based, and data-driven methods. However, these methods cannot achieve fully achieve effective analysis of the real-time effects of multi-factor loads on bridge deformation, particularly in the presence of dynamic and simultaneous loads such as wind or temperature variations. A novel data-driven method is proposed based on a state-of-the-art real-time updating artificial neural networks (ANNs) algorithm to investigate the real-time coupling relationship between multi-loads and bridge deformation, enabling real-time prediction of bridge deformations. Additionally, the real-time characteristics between structural deformation and multi-loads are explained by incorporating SHapley Additive exPlanation (SHAP) in harsh natural environments. The proposed method has been validated on the 1,006-meter Forth Bridge in Scotland, showing high accuracy in real-time displacement prediction. The 9-day testing dataset demonstrated the R² values for Y and Z direction deformations were found to be 0.98 and 0.87, respectively. The performance metrics for each day indicated that the majority of Y and Z direction deformations had R² values exceeding 0.8, with RMSE and MAE values below 30 mm. The SHAP analysis revealed that an increase in wind speed leads to intensified Y direction deformation (larger SHAP values), while temperature has a significant impact on Z direction deformation (smaller SHAP values). Moreover, the weight influences of each load on the deformation are not fixed. The study's findings demonstrate that the proposed method enables accurate long-term prediction and assessment, allowing precise monitoring and prevention of abnormal risks in bridges under harsh environmental conditions.