Rapid seismic damage state prediction of the subway station structure using the pre-trained network and convolutional neural network

A rapid prediction model for the seismic damage state of subway station structures based on the pre-trained network (PTN) and convolutional neural network (CNN) is proposed. The model directly maps the ground motions to the structural damage states. Firstly, 512 stochastic ground motions are generated, and the nonlinear time history analysis (NLTHA) is conducted. The 1D ground motion is converted into the 2D image format and labeled. Then, the image augmentation technique balances and expands the dataset. Finally, performance and efficiency tests are conducted on the PTN-based deep learning model (PTN-DLM) and CNN-based deep learning model (CNN-DLM). The results indicate that the prediction accuracy of the PTN-DLM is as high as 94.57 %, with 2.35 minutes. The prediction accuracy of the CNN-DLM is 82.60%, with 0.33 minutes. The NLTHA performed by the finite element model is considered to be 100% accurate and takes approximately 240 minutes. Therefore, the proposed PTN-DLM is the most cost-effective.