ML Modelling on strength reduction of fire-damaged RC column via numerical data

Abstract

In this thesis, a dataset for strength reduction of fire-exposed RC column which is described in terms of P-M diagram downsize is built, and a Machine Learning (ML) based RC members mechanic analysis is proposed by making a model which predicts P-M reduction as fire exposure lasts. Since a fire exposed RC member experience physical property changes and non-mechanical deformation due to temperature increase and chemical reactions in it, it shows different mechanical behavior to the RC member before fire exposure. A dataset is consisted of numerical analysis of fire exposed RC column P-M diagram on sample set of 1770 RC section with different width, height, and reinforecd steel conditions. With this P-M diagram data, a model is built which predicts the P-M diagram reduction ratio given section input using several ML algorithms, kernel SVM, ANN, Random Forest, XGB and LGBM and each are compared. This model achieved stable result with Mean Absolute Percentage Error (MAPE) under 2%.