UNDERSTANDING MEM-MODELS AND INTERPRETABLE MACHINE LEARNING USING EXPERIMENTAL LABORATORY DATA THROUGH PROGRAMMING

Abstract

Structural damage and failure are of particular concern in engineering because of their severe human and economic impacts. Apart from design and finite element method, the use of experimental laboratory and site data to study damage and failure has been vital and has made great strides over the last few decades. In this study, most of the effort is devoted to the first topic, which is to study the datasets of three wood shear wall cyclic responses involving damage. In addition, the author supports the second topic, which is the prediction of shear failure in prestressed concrete girders using a shear database. The goal of this study is to understand nonlinear dynamics, especially the hysteresis of wood shear walls, and nonlinear methods involved in both the topics.The first nonlinear method is for hysteresis modeling, named after mem-modeling. The second nonlinear method is interpretable machine learning (IML). Both methods are advanced; however, programming was exercised in this study to catalyze the learning process. An existing code for mem-modeling was adapted and modified to analyze the hysteresis of the wall specimens. The loading protocol is different and the physical process is complicated, which imposed significant challenges in adapting the existing code to process the data. Two remedies were proposed with moderate success. Limitations and future work are identified. An obsolete interpretable machine learning MATLAB® code was upgraded. The author created a Python® script for interpretable machine learning with the same functionality as an existing MATLAB® counterpart. With these advancements, this study has made a contribution to the development of the two advanced methods.