Prediction of rocking behaviour via ensemble machine learning methodologies

Abstract

This thesis demonstrates the application of machine learning methods to predict and interpret the response of rocking rigid structures when subjected to recorded ground motions. This work adopts random forest algorithms to predict the rocking response of a rigid block. Specifically it focuses on two problems, the classification, and the regression problem. The classification problem aims to efficiently predict the rocking class: whether a block, after commencing rocking motion, overturns or undergoes safe rocking motion. The regression problem aims to predict the peak rocking response in the case of safe-rocking only. The study achieves this by training a machine learning model, namely, the classification and regression tree random forest algorithm to predict the rocking class, and peak rocking amplitude for a specific rigid block and recorded strong ground motion, respectively. The input data used are dimensionless intensity parameters, which are parameters that unite characteristics of the rigid block and the strong ground motion. Importantly, this thesis sheds light on the "black box" behaviour of machine learning methodologies, and therefore, utilises various "interpretable machine" learning tools to highlight the importance and interactions between strong ground motion parameters. Specifically, this thesis employs Minimum Redundancy Maximum Relevance to identify important dimensionless intensity measures for both the classification and regression problem. This analysis yields intensity measure vectors for the classification and regression problems, which contain efficient dimensionless intensity measures that provide predictive accuracy on par with their black-box counterparts. In addition, it adopts Shapley additive explanations and partial dependence plots to interpret the predictions and reveal hidden trends of the nonlinear rocking response. This work highlights the importance of velocity characteristics of the seismic signal on the overturning mode of a rocking block. Further, it shows the importance of velocity, duration, and potentially energy characteristics on rocking amplification. Importantly, it reveals, for the first time in rocking literature, hidden trends in the correlation between rocking demand and seismic characteristics through visual representations, and critical values of each dimensionless intensity measures that promote overturning.

Date of Award	2024
Original language	English
Awarding Institution	The Hong Kong University of Science and Technology
Supervisor	Ilias DIMITRAKOPOULOS (Supervisor)