Seismic response analysis of subway station structure under random excitation based on deep learning and PDEM

The randomness of ground motions and soil properties has a significant impact on the seismic response of underground structures. The probability density evolution method (PDEM) is a powerful stochastic dynamic analysis method. By decoupling the physical space and probability space, any number of random excitation variables can be incorporated into the dynamic system, thereby accurately capturing the time-varying evolution of structural dynamic response. However, due to the expensive computational cost of nonlinear time history analysis (NLTHA) of underground structures, the application of the PDEM that requires many calculation samples is severely limited. In view of this, this paper proposes a lightweight stochastic seismic response analysis method for the subway station structure based on deep learning and PDEM. Firstly, mathematical models of stochastic ground motions and random soil shear wave velocity profiles are constructed. Two types of random excitations are coupled using the high-dimensional space optimization point selection method. The one-dimensional convolutional neural network (1D-CNN) is used as a proxy model to map the free field's nonlinear time-history response to the subway station structure's nonlinear time-history response. The computational cost of stochastic dynamics analysis has been significantly reduced. Subsequently, the predicted results are combined with PDEM to conduct a comprehensive analysis and evaluation of the subway station structure's seismic response and seismic performance. The research results indicate that the proposed 1D-CNN model has superior predictive performance, with a computational cost of approximately 92.5% less than the traditional process. The evolution of the subway station structure's inter-storey drift ratio (IDR) is mainly influenced by stochastic ground motions. However, under the combined random excitation, the peak value of IDR is larger and its occurrence time is earlier, so the establishment of earthquake evacuation plans should refer to this working condition. Both the randomness of soil properties and the randomness of ground motions need to be considered, otherwise, underground structures’ seismic performance will be overestimated. This study provides new technologies and directions for seismic analysis and design of underground structures. © 2024 Elsevier Ltd.