Seismic responses of tall pier bridges are usually estimated with nonlinear time history analysis (NLTHA) since it is able to provide rigorous results while the time consumption is acceptable with the improvement of computers. Note that parallel computing employing multiple computers might be required to facilitate estimating the performance of numerous bridges in highway networks after earthquakes. Recently, deep learning techniques have been recognized as promising alternatives for predicting structural responses in earthquake engineering with significantly improved time efficiency. Therefore, this paper develops a fast seismic performance estimation methodology using deep learning procedures to rapidly predict the seismic demands of tall pier bridges. The efficiency of the employed techniques is verified through illustrative examples, by comparing the predicted responses with those obtained from NLTHA under several types of input motions. The results show that when trained following appropriate steps, the deep learning models could provide satisfactory prediction for shear force, bending moment, as well as section curvature ductility. Additionally, the time efficiency of deep learning models is shown increased by about 97% compared with NLTHA, which might be further improved for more complex structural systems. Further parametric analysis reveals that the efficiency of selecting proper input variables for deep learning models could be significantly improved by considering the physical characteristics of structures; e.g., structural dynamic properties and interaction between structure and ground motion. This methodology is believed especially favored evaluating the seismic performance/post-earthquake resilience of highway networks containing thousands of bridges, in which conducting NLTHA for each bridge is prohibitively computational demanding and might delay rescue operations.
Read full abstract