A Machine Learning-based Study On Post-Earthquake Damage Assessment Model And Ground Motion Characterization Method For Buildings

Seismic hazards usually cause serious economic losses and a large number of casualties,which bring incalculable losses to the economy and society.An effective pre-earthquake predic-tion and timely&accurate post-earthquake damage assessment can significantly reduce the damages caused by the disaster events,which are important approaches to enhance the disaster prevention and mitigation capabilities and to control the disaster damages.There have been a lot of studies focusing on traditional physics-based methods,while techniques based on machine learning（especially deep learning）show great potential in terms of balancing accuracy and computational efficiency,and making comprehensive use of the existing data accumulation.In this paper,a machine learning approach is adopted to carry out an in-depth study on the problems of post-earthquake damage assessment for buildings and characterization of ground motion,as follows:（1）A machine-learning based fast seismic damage assessment and prediction method is proposed which considers the influence of multivariate ground motion intensity index and structural feature parameters on the ground vibration damage for buildings,and compares and selects the input parameters by an iterative embedded method.Subsequently,a case analysis for building post-earthquake damage assessment and damage index prediction has been carried out on a subset of optimal input features to validate the efficiency and accuracy of the method.Finally,the predictive model has been interpreted using the shapley additive explanations method,and the interaction effects of input features have been analyzed.（2）In response to a wide variety of scenarios for building post-earthquake damage assessment,considering different ground motion feature characterization methods and neural network algorithms,it is proposed a ground motion characterization method based on Mel-frequency cepstral coefficients,and an adaptive ground motion characterization network module is designed.This module improves the prediction accuracy of the post-earthquake damage assessment network for buildings by identifying ground motion features that are closely related to the behavior and response of the structure under dynamic loading.Then,a dataset is constructed based on a typical building and the designed neural network,including a baseline model M₀ and three comparison models（M₁,M₂ and M₃）,are systematically analyzed.The results show that the three comparison models demonstrate greater adaptability and faster computational efficiency for the post-earthquake damage assessment task compared to the baseline model.（3）With respect to the complexity of the time-frequency characteristics for ground motion,three network modules for ground motion characterization are designed based on the wavelet transform theory,i.e.,the discrete wavelet transform module,the static discrete wavelet transform module,and the continuous wavelet transform module.The accuracy and efficiency of the method are validated by testing on the dataset,where the static discrete wavelet transform module shows the best performance.In addition,compared to the hand-crafted ground motion features,there are more significant changes in the ground motion features extracted by both discrete wavelet transform module and static discrete wavelet transform module after training,which contain more information about the post-earthquake damage of the building（4）From the perspective of multi-domain information fusion,a fusion framework based on a one-dimensional convolutional neural network and three feature fusion methods（end fusion,intermediate fusion,and double fusion）are designed to effectively improve the feature extraction and nonlinear fitting ability for building seismic damage prediction models.Then,an ablation experiment is implemented to validate the effectiveness of the fusion methods,and the accuracy and computational efficiency of the fusion network.（5）In terms of the seismic damage assessment,both the designed building seismic damage prediction network and the ground motion characterization module are tested.Moreover,the reliability and generalization performance of the proposed network model,as well as the applicability of the ground motion characterization method,are validated by testing on different buildings.Machine learning-based post-earthquake damage assessment model and ground motion characterization method for buildings are proposed in this paper,which obtains more satisfac-tory results under different experimental conditions and shows better generalization perfor-mance in the testing of cases,thus it has good application potential.