Hybrid machine learning models for classifying failure modes of unstiffened steel plate girders subjected to patch loading

Abstract

This paper, for the first time, develops novel hybrid machine learning models that combine Support Vector Machine, Naïve Bayes, K-Nearest Neighbors, Decision Tree, Random Forest, Adaptive Boosting, Gradient Boosting, and Extreme Gradient Boosting, with Moth-Flame Optimization algorithm, called MFO-SVM, MFO-NB, MFO-KNN, MFO-DT, MFO-RF, MFO-AB, MFO-GB, and MFO-XGB, respectively, to classify the failure modes of unstiffened steel plate girders subjected to patch loading. The experimental database is collected from the tests reported in the literature to develop the hybrid machine learning models. The Synthetic Minority Over-sampling Technique is adopted to address the imbalanced class problems of the database. Hyperparameters of machine learning models are optimized using the Moth-Flame Optimization algorithm. The performance of hybrid machine learning models is evaluated and compared through the confusion matrix using several metrics. The results show that the MFO-RF model proves more effective than other models in identifying the failure models of unstiffened steel plate girders. The Shapely Additive Explanations method is used to analyze the contribution of each variable to the failure modes of unstiffened steel plate girders both globally and locally. Web thickness is the most influencing feature for identifying the failure modes. The MFO-RF model is finally utilized to develop a web application for practical use at low cost and effort.