Data-driven Dynamic-classifiers-based Seismic Failure Mode Detection of Deep Steel W-shape Columns

Abstract

It is vital to assess the health of buildings following a major earthquake. New technologies such as deep learning algorithms have grown increasingly tempting in such rapid applications because of their increased reliabilities and simplicity to traditional methods. Due to the kinematics of steel moment frames, inelastic deformations tend to concentrate within the steel column during an earthquake, resulting in local or global buckling. Rapid failure mode detection of the existing deep steel W-shape columns (DSWCs) cannot be quickly identified due to a lack of comprehensive empirical and mechanics-based models. This research proposed a machine learning (ML) algorithm based on the state-of-the-art techniques of dynamic classifiers for failure mode forecasting of the DSWCs using an experimental database and illustrated why the ML model suggests a specific failure mode for a particular sample. The database was created by combining 939 instances from various studies that have been published. A total of six machine learning models based on Dynamic Selection strategy were implemented. Three metrics, i.e., accuracy, precision, and recall, were used to evaluate the performance of models. As a result of the extensive examination, a machine learning model based on the META-DES model was proposed. In the training stage, Overall Local Accuracy, A-Priori, and META-DES algorithms, received the highest score (>0.96) across all criteria. The META-DES model correctly predicted the failure mode of the DSWCs with an accuracy of 0.907 in the testing phase. The META-DES algorithm performed better than previous methods which are employed to identify the failure mode.