Research On Earthquake Damage Prediction Based On Machine Learning

Accurate assessment of earthquake damage is of vital importance for pre-earthquake disaster prevention and mitigation,post-earthquake disaster relief and rapid reconstruction.Most of the existing studies based on actual earthquake damage assessment have too little sample data,limited to a specific region and a certain structure type,and the number of data samples used is also limited.Machine learning can learn complex nonlinear relationships between parameters by inputting large data samples as training.Compared with traditional methods,machine learning methods are more computationally efficient,more scientific,and easier to apply.To this end,this paper designs a method that integrates multidimensional information such as ground shaking information,building attribute information and site information based on 378,037 actual building damage data from the Great East Japan Earthquake of March 11,2011,using the random forest algorithm for predicting the degree of earthquake damage to buildings.The main work is as follows:(1)Firstly,we compiled 378,037 building damage data from the March 11,2011 Great East Japan Earthquake in detail,built a database of earthquake damage information by data matching and data pre-processing,and selected 12 influencing factors such as peak ground acceleration,spectral acceleration,number of stories,and structure type as model inputs,and accurately predicted the damage caused by earthquake damage to buildings using four damage states of earthquake damage levels(minor damage,moderate damage,severe damage,and collapse)released by he American Applied Technical Council(ATC-13),giving full play to the computational characteristics of machine learning efficiency.(2)Four classical machine learning algorithms,Random Forest,Support Vector Machine,XGBoost and Decision Tree,are introduced in detail to compare the prediction accuracy rates of the four methods,conclude that the classification performance predicted by the random forest method is more dominant,and analyze the applicability of the random forest method in earthquake damage prediction research.(3)The classification performance of the model is further investigated based on the Random Forest approach,using both data balancing process and model hyperparameter optimisation.The results show that:undersampling to achieve sample balancing by reducing the number of samples in most classes in classification causes the model to lose a large amount of important information,and the model accuracy is better with the SMOTE method using oversampling;the model performance given by the hyperparameter optimisation method using Bayesian optimisation is higher than that given by the learning rate curve.(4)In order to improve the applicability of the model and the performance of the model in post-earthquake decision making,a prediction target optimization method was proposed,using the FEMA-356 specification to classify the prediction targets into two application performances of immediate occupancy and life safety,simplifying the original four classifications into two classifications,and the results showed that the model accuracy was improved to 73.0%and 87.5%,respectively.(5)The prediction performance of the model optimised for the prediction objectives is verified,including the following three aspects:① Taking the Fukushima Prefecture area in Japan as an example,the loss distribution of each damaged building is annotated using the QGIS GIS tool to verify the spatial distribution of the model in terms of the generalisation of the spatial distribution of the buildings,and the results show that the model prediction is basically in line with the real distribution.②1650 buildings damaged by the Mw6.3 magnitude earthquake that occurred on January 20,2014 in the New Zealand region were used as a single seismic event for validation,and the results showed an overall accuracy of 79.9%under the life safety performance and 73.0%under the immediate occupancy performance.③Using 137,648 buildings damaged by earthquakes that occurred in the New Zealand region from 2003 to 2014 as multiple seismic events for validation,the results showed an overall accuracy of 73.3%under life safety performance and 70.3%under immediate occupancy performance.