Sensitivity Analysis Of Seismic Parameters Based On Machine Learning

The accurate assessment of seismic damage at regional level is important for the strategies of disaster prevention and mitigation and the construction of resilient cities.The mapping relationship between structural parameters,ground motion parameters and structural damage index is established based on machine learning method,which can effectively balance computational accuracy and efficiency,thus it is an effective method for regional seismic damage assessment.As an important index to describe the uncertainty of ground motion,seismic parameters are both the key to determine the potential impact on structural damage,and the premise of establishing the matching relationship between real ground motion and seismic damage database.Therefore,the purpose of this paper is to determine the key seismic parameters through sensitivity analysis,so as to improve the efficiency of seismic damage assessment and reduce the sample demand of the database.The research route of this paper is as follows:firstly,the commonly used ground motion intensity indexes are collected,and the dimension reduction is carried out based on the correlation analysis.Then,the sensitivity analysis of SDOF(Single Degree of Freedom)and MDOF(Multiple Degrees of Freedom)structures is carried out to obtain the key parameters.Finally,the information integrity of key parameters is verified through the structure response prediction of typical frame structures at macro structure level and component level.The main contents and conclusions are as follows:(1)The dimension reduction is carried out through the correlation analysis based on far-field ground motions of typeⅡsite.The selected ground motion record is type-Ⅱsite and far-site motion,and a total of 734ground motions are obtained after outlier elimination and K-Means clustering.55 ground motion parameters proposed are collected,and the batch calculation of seismic parameters is completed based on the selected ground motion records.30 SDOF structures are constructed based on Open Sees,and batch elastoplastic time history analysis is carried out.Based on the linear correlation analysis between the ground motion parameters and the logarithmic linear correlation analysis with the maximum displacement response of SDOF,the dimension of the ground motion parameters is reduced to 11 parameters.(2)The structural response prediction algorithms of SDOF and MDOF are established based on the extra trees regressor.Taking five commonly used ensemble methods as alternative models,the model selection is completed based on multiple SDOF sample data,and the extra trees regressor is determined.Taking 11 ground motion parameters and structural basic periods as input features,the maximum responses of SDOF and MDOF structures as output features,a response prediction model is constructed.The algorithm turns to be a stable and reliable model,as the R~2determination coefficient of the testing set reaches 0.80.In addition,the SDOF response is the maximum displacement,and the MDOF response includes the maximum top displacement,the maximum inter-story drift ratio and the maximum base shear force.(3)The sensitivity analysis of seismic parameters is carried out based on SDOF and MDOF structural response prediction model.The sensitivity analysis of input features is carried out based on the calculation of feature importance coefficients in extra trees regressor.Results show that the sensitivity of ground motion parameters is strongly related to the structure period.The highest correlated parameters turn to be Mackie ground motion intensity ASI in the short period range,Riddell index Iv in the middle period range,the peak value of displacement spectrum Sd＿max and the maximum incremental displacement MID in the long period range.Sample size analysis in conducted with the above four key parameters as input features,and it turns out that the sample demand is only 250 with the accuracy reaching 75%.It is proved that the key seismic parameters already contain most information of ground motion records.(4)The results of sensitivity analysis are verified at the macro-structure level and element level of typical frame structures.The elastic design of typical frame structures of 3,6 and 9 stories is completed by YJK software,and batch elastoplastic time history analysis is carried out based on Open Sees.Then the structural damage index is extracted,that is,the maximum inter-story drift ratio and the maximum base shear force,and the element deformation index,that is,the proportion data of beams and columns with different performance states in the whole building.The adaptability of the above conclusion is verified by the construction of the prediction model,sensitivity analysis and sample size analysis of the structural response.In addition,a classifier is constructed with the key ground motion parameters as input features and element deformation index as output features.The accuracy of the classifier reaches 70%for beams and columns in performance states 1 and 2,thus the information completeness of the key seismic parameters is verified.