Seismic Damage Analysis Of Building Components Based On Machine Learning Algorithms

The economic losses caused by earthquakes are enormous,so how to minimize the losses caused by earthquakes is an issue that is particularly worth thinking about.Based on the experience of New Zealand and other countries in responding to earthquakes,the establishment of a mature and complete catastrophe risk model for China is one of the most effective ways to reduce this loss.This is because the establishment of an earthquake catastrophe risk model can not only effectively reduce the losses caused by earthquakes,but also gradually form an earthquake insurance database belonging to China,which will lay a good foundation for the rapid analysis of earthquake risks in the future.The seismic vulnerability of buildings is a crucial part of the earthquake catastrophe risk model,and the damage of buildings often starts from the components,so the vulnerability analysis of components is more helpful to clearly understand the risk.Through reading a large amount of literature,it can be seen that the traditional method of analyzing the vulnerability of building components is often for a single structural type,usually does not take into account the uncertainty of the building as a whole and the importance of nonstructural components,and is often subject to cost constraints,etc.That’s why the conclusions drawn are often not universal.The development of science and technology,the emergence of big data and machine learning algorithms as well as the continuous updating and improvement of earthquake damage data provide a powerful research platform for the study of component vulnerability.In this paper,based on the detailed seismic data,damage data,and claim data in the database of New Zealand,this paper investigates the contribution of structural components(e.g.,wall frames,piles)and non-structural components(wall cloths,etc.)to the seismic loss of a building by establishing a machine learning model of the seismic loss of components from the building components themselves,it integrates the XGBoost and SHAP methods to validate the importance ranking results obtained by using the Random Forest model to validate the importance ranking results,finally,the effect of liquefaction on the susceptibility of building components is analyzed.This paper mainly utilizes machine learning algorithms to carry out the following work:1.First,the background and research significance of the topic selected for this thesis are discussed.Then,the traditional component vulnerability analysis techniques used domestically and internationally are summarized.Finally,the innovation and viability of using machine learning techniques for component loss analysis are examined.2.Initially,The New Zealand earthquake hazard characteristics and earthquake insurance database were described in detail,and a total of 12 types of components used for significance analysis were compiled in accordance with the national building component classification standard,and the details of the classification of the different component types were explained.The 12 loss component datasets were then classified into four labeling categories according to the American(ATC-13)damage standard,and this sample set was applied to a random forest model.3.machine learning algorithms and various importance analysis methods that can help to complete this study are presented.The random forest algorithm is used,and the seismic component losses are evaluated to determine the importance ranking of the various components affecting the economic losses of the earthquake.The training set is tuned by adjusting the hyperparameters,and the training set is assigned to obtain a more optimized random forest model.4.The results obtained from each of the FI method,PI method and sensitivity analysis method are combined and analysed by applying the normalisation method and the summation of assignment ranking method to derive the importance ranking of component losses based on the random forest model.The importance ranking of the loss components is also discussed from the perspective of liquefaction and non-liquefaction,taking into account the liquefaction phenomenon in the New Zealand earthquake.Finally,the importance ranking results obtained from the random forest model are compared and validated by integrating the XGBoost and SHAP methods in machine learning and existing academic results.