Rapid seismic damage state assessment of infilled RC frames using machine learning methods
Infilled reinforced concrete (RC) frame structures are one of the most common structures.It is found that infilled walls have a significant impact on seismic performance of RC frames in past earthquake damage investigations and experimental tests.To accurately and rapidly assess seismic damage states of infilled RC frames after an earthquake,660 infilled RC frames were firstly designed based on different building structure information (i.e.the seismic design intensity,constructed period,number of stories,story height,number of bays and the filling rate),then the non-linear time history analysis was performed for the 660 infilled RC frames with 10 ground motions in OpenSees.6600 data points were gained from the analysis,resulting in a dataset which was used to develop seismic damage state assessment models of infilled RC frames.Based on the dataset,nine machine learning models predicting seismic damage states of infilled RC frames were developed using naive Bayes (NB),K-nearest neighbors (KNN),decision tree (DT),artificial neural network (ANN),random forest (RF),adaptive boosting (AdaBoost ),extreme gradient boosting (XGBoost ),light gradient boosting machine (LightGBM ),category boosting (CatBoost ) algorithms.The results indicated that CatBoost and RF models had the highest prediction accuracy for the seismic damage state which was 0.93 in testing dataset,followed by LightGBM and XGBoost models with an accuracy of exceeding 0.90.Compared with actual damage investigated in the past earthquakes indicating that RF and CatBoost models achieved an identical accuracy of 47％.However,the difference in the remain damage states within one damage state level occupied 76％ for CatBoost model,which was higher than that of RF model.Based on the CatBoost,importance analysis was performed for different input variables.It is found that three input variables had the greatest impact on infilled RC frame,including seismic design intensity (SDI),peak ground velocity (PGV) and the spectral acceleration at Sa(0.4 s).Furthermore,the importance of the number of stories on the seismic damage state for infilled RC frames increased as the increase of the number of stories.
infilled RC framesmachine learningdamage statedamage assessmentfinite element model
万方数据
维普
