Classification method for failure modes of RC columns based on class-imbalanced datasets. (February 2023)

• Record Type: Journal Article
• Authors: Yu, Bo; Xie, Longlong; Yu, Zecheng; Cheng, Hao
• Is Part Of: Structures. Volume 48(2023)
• Page Start: 694
• Page End: 705
• Publication Date: 2023-02
• Subjects: Reinforced concrete columns -- Classification of failure modes -- Class-imbalanced datasets -- Synthetic minority over-sampling -- Tomek links -- Gradient boosting decision tree; Structural engineering -- Periodicals; 624.1
• DOI: 10.1016/j.istruc.2022.12.063
• Languages: English
• ISSNs: 2352-0124

Abstract

Abstract: In order to overcome the limitation that traditional machine learning (ML) techniques cannot accurately classify the failure modes of reinforced concrete (RC) columns due to the imbalance class distribution in datasets, an efficient classification method for four types of failure modes of RC columns based on class-imbalanced datasets was proposed. A new treatment method for class-imbalanced datasets was proposed first by combining the synthetic minority over-sampling (SMOTE) with the Tomek Links technique. Then an efficient classification method for four types of failure modes including flexure failure (F), flexure-shear failure (FS), shear failure (S) and splitting failure (SF) for RC columns was developed based on the gradient boosting decision tree (GBDT) algorithm. Finally, the proposed method was validated by comparing with untreated method and four traditional treatment methods for class-imbalanced datasets and six typical machine learning methods based on a total of 423 sets of experimental data for RC columns (including 253 sets of F, 65 sets of FS, 53 sets of S and 52 sets of SF). The results show that F1 scores and Kappa coefficients of the minority classes (e.g., FS, S and SF) were increased about 12% and 0.14 respectively when comparing with untreated method and four traditional treatment methods for class-imbalanced datasets, while F1 scores and Kappa coefficients of the minority classes were increased about 19% and 0.17 respectively when comparing with six typical machine learning methods. The proposed class-imbalance treatment method is a hybrid sampling method, which overcomes the limitation of both over-sampling and under-sampling techniques.