Excessive deflection which exceeded expectations has been identified as one of the main problems contributing to long-term performance degradation of long-span girder bridges. Although the 3D refined finite element model of the full bridge can improve the prediction accuracy of long-term creep deformation, there is still a lack of efficient creep-sensitive structural judgment methods in the preliminary design stage. This study was thus intended to develop a new evaluation method and model to identify the possibility of excessive deflection will occur by utilizing simple structural design parameters, which contributed to new understandings of the creep deformation of bridge structural level. A finite element model database established by 20 long-span prestressed concrete girder bridges with actual project backgrounds was used for providing training, testing, and validation sample datasets for the artificial neural networks (ANN). The Back Propagation (BP) model was chosen for creating mappings among variables and results and solving complex problems with a fairly accurate prediction even without sufficient information. The indicator of deflection-span ratio difference between creep models was proposed as the key analysis factor. A case study was conducted to validate the effectiveness of the ANN model. The mean absolute error (MAE) between the results of the ANN model and FE analysis was 6.25%, which indicated a high consistency. The analysis of bridge design parameters using the ANN model revealed that it had a high risk of excessive deformation, which was consistent with the conclusion of the actual situation. Multiparameter sensitivity analysis results from the model also suggest that the side-span ratio, the main span, and the deflection caused by the pavement are the most three significant impacts compared to different input variables in this model. Based on the results, it can be concluded that the performance of the ANN-based technique can predict possible excessive creep problems in advance.
Read full abstract