Machine learning models for predicting maximum displacement of triple pendulum isolation systems

Abstract

Maximum displacement is an important engineering demand of an isolation system, including systems using triple friction pendulum bearings, during earthquakes. This response can be accurately predicted by time-history dynamic analysis of the nonlinear model of the system. However, this analysis approach is time-consuming and requires skillful analysts. To remedy the cumbersomeness, this study developed four machine learning models to confidently predict the important demand using limited number of parameters defining isolation system and earthquake event. Specifically, random forest, gradient boosting regression tree, adaptive boosting, and extreme gradient boosting approaches were employed to develop the machine learning models. The input features to the models include eight constitutive parameters of the triple pendulum bearings in the isolation system and five spectral accelerations at control periods of the average spectrum of the site. The database for constructing the machine learning models was obtained from time-history analysis of lumped-mass nonlinear model of isolation systems subjected to earthquake ground motions. The performance investigation showed that all proposed machine learning models can confidently predict the maximum displacement from the time-history analysis procedure. Among the four models, extreme gradient boosting model possesses the highest accuracy with an average ratio between analysis and predicted values of 0.9999 and a coefficient of variation of 0.017. A graphical user interface module based on this machine learning model was developed for practical uses. The module was written in Python and is free for download at GitHub.