Search

The impact-echo (IE) method is effective for evaluating invisible defects. However, it might return misleading results when its signals are invalid. This challenge aggravates when the tests are conducted using robotic devices that automatically collect massive data. This study proposes an automatic method to eliminate invalid signals based on the ResNet model. First, the signals are visualized into two-dimensional images as the input for ResNet. The input data can then be classified into valid and invalid data via the ResNet model, which is trained with 11,290 signals and tested with 5664 signals. Finally, defects can be detected using the dominant frequencies of the valid-class data. A case study with IE data from two concrete bridges was employed to validate the feasibility of the proposed approach. The results indicate that the method can achieve an average accuracy of 90.6% for eliminating invalid signals and significantly improve the IE test accuracy.

The deformation of long-span suspension bridges in multiple loads is an important indictor to reflect their operation state. However, the correlation between multiple loads and structural deformation is difficult to quantify. Therefore, this study proposes an explainable machine learning model for the load-deformation correlation in long-span suspension bridges using eXtreme Gradient Boosting (XGBoost) and SHapley Additive exPlanations (SHAP). Firstly, the structural health monitoring system for a suspension bridge was used to construct the dataset for the training and testing of XGBoost model. Herein, temperature, wind and vehicle loads were used as the input variables, while midspan deflections and expansion joint displacements were treated as outputs. Subsequently, the hyperparameters of XGBoost model were optimized using grid search and 5-fold cross-validation to ensure its prediction performance. Then, the prediction results were compared with other four machine learning methods (i.e., linear regression, artificial neural networks, gradient boosted decision trees and CatBoost). Finally, the correlation between different loads and displacement responses were explained by the SHAP method to identify the contribution of the loads on deformation. The results show that the XGBoost model has the highest prediction accuracy. Compared to vehicle and wind loads, temperature significantly affects the deformation of long-span suspension bridges during daily operation. The effects of temperature and wind on bridge deformation are independent, and there is no significant interaction between these two factors.

The Tian'e Longtan Bridge, with a span of 600m, is the world's longest arch bridge, surpassing similar structures by 34.8%. To ensure the construction safety of Tian'e Longtan Bridge, a large-scale model test was conducted using the stress equivalence principle, with a scale ratio of 1:10. The experiment simulates the entire construction process of gradually wrapping concrete by layers and segments on a stiff skeleton, with data collected from 401 stress points and 28 displacement points, revealing the evolution of structural deformation and stress during arching process. The results show that during the pouring of encasing concrete, the maximum compressive stress in the steel tubes is 235.5MPa, located at the upper chord of the midspan, while that of the encasing concrete is 10.6MPa and located at the bottom plate of the 3/8 span. The maximum stress ratios that cannot be transmitted between the concrete on both sides of the interface between layers and segments are 6.8% and 5.8%, respectively, which demonstrates the reliable stress transmission at the interfaces. The elastic deflection at midspan is 53.2mm when the bottom slab is closed, while only an additional deflection of 3.3mm occurs during subsequent construction of the web slab and top slab, which indicates that most deflection occurs before the closure of the bottom plate. The deflection at the arch crown caused by shrinkage and creep is 42.3mm. In contrast, the calculation result based on the uniaxial shrinkage and creep theory is 54.7mm, 29.3% larger than the test result.