Research On Estimation Method For Seismic Fragility Of Structures Based On Unsupervised Machine Learning

The estimation of structural fragility plays a fundamental role in performance-based earthquake engineering(PBEE)framework.As an effective procedure for estimating structural fragility especially collapse fragility,incremental dynamic analysis requires incrementally increasing an intensity measure corresponding to a scale factor on a ground motion set and a large number of nonlinear time-history analyses,and thus may result in low computational efficiency.How to estimate structural fragility accurately and efficiently has become a pressing issue in the assessment of seismic performance and risk of structure.Although there have been many studies on methods for computationally efficient fragility estimation of structure,these methods cannot strike a balance between accuracy and computational efficiency in the estimation of structural fragility.As an effective data mining method that can directly use unlabeled data for analysis,unsupervised machine learning has recently been used in various fields.Introducing unsupervised machine learning algorithm into the estimation of structural fragility can effectively solve the problem of the contradiction between accuracy and computational efficiency in the seismic fragility estimation.In light of this,this study aims to propose an unsupervised-machine-learning-based method for the seismic fragility estimation when considering accuracy and computational efficiency.The main contents of this study are as follows:(1)The ranking of ground motion parameters used for structural seismic capacity prediction is established and the representative ground motion parameters are selected based on ranking results.A total of 36 ground motion parameters are selected in this study as input variables,and the seismic capacity of a generic set of single-degree of freedom(SDOF)systems are used as output variables to establish regression model.The ranking of ground motion parameters is established in terms of sensitivity and frequency obtained from results of the linear regression model on various SDOF systems.The representative ground motion parameters,Saavg,FIV3,FR1,FR2,CAV,DSI,PGV,td,Pa,that are less affected by various ground motions,structural types,limit-states and various regularizers are selected based on the ranking results.(2)The critical features of ground motion parameters are extracted by using the feature extraction technique and a compound-ground-motion-parameter developed with a combination of critical features is used for characterizing the potential structural damage of ground motion.The most frequently used feature extraction techniques,namely,Exploratory factor analysis,Kernel principal component analysis,Locally linear embedding,Isometric mapping,Multidimensional scaling,t-distributed stochastic neighbor embedding,are summarized and basic principles of these feature extraction techniques are introduced.The critical features of ground motion parameters are extracted by using EFA,a compound-ground-motion-parameter,IMEFA,developed with a combination of critical features is used for characterizing the potential structural damage of ground motion.The correlation between IMEFA and damage measures was investigated and results are statistically organized to evaluate the influence of various ground motions,structural types and yield strength coefficient.(3)A method for efficient estimation of structural fragility based on critical features clustering is proposed.In this method,the most frequently used feature extraction techniques are used to preferentially extract critical features of ground motion parameters,which are applied to construct the feature space.Then,the Bisecting Kmeans clustering algorithm is applied to the feature space to obtain a subset of ground motions from the entire ground motions set until the fragility curve converges.Based on analysis results from a generic set of SDOF systems,the effects of the number of critical features of ground motion parameters and convergence tolerances on the performance of the proposed method are investigated,and the optimal number of critical features of ground motion parameters and convergence tolerances are given.The efficiency of the proposed method is successfully demonstrated through five reinforced concrete(RC)momentresisting frame buildings and one steel moment-resisting frame building.(4)A method for efficient collapse fragility estimation of structure based on spectral acceleration clustering is proposed.In developing the method,the Bisecting Kmeans clustering algorithm is applied to the feature space that is constructed;the construction is done by using the spectral acceleration over a certain period interval to obtain a small number of ground motions from the entire ground motions set until the collapse fragility curve converges.Based on analysis results from a generic set of SDOF systems,the effects of the period ranges for spectral acceleration and convergence tolerances on the performance of the proposed method are discussed,and optimal period ranges for spectral acceleration and convergence tolerances are given.The efficiency of the proposed method is successfully demonstrated through six moment-resisting frame buildings.