Life-cycle corrosion behaviors and failure modes prediction for chloride-induced corroded RC columns based on explainable generalized additive model
Year: 2025
Type: article
Abstract: In order to overcome the disadvantages of class imbalance at the data level and the 'black box' at the algorithm level of traditional machine learning (ML) models for failure modes of chloride-induced corroded reinforced concrete columns (CRCC), an innovative explainable generalized additive (EGA) model has been developed using the improved hybrid sampling technique and the enhanced explainable boosting classifier. An experimental database consisting of 230 specimens for failure modes of CRCC with rectangular sections was established first. Then an improved hybrid sampling technique for failure modes of CRCC was proposed by combining the advantages of oversampling and undersampling techniques. The generalization ability of the EGA model was confirmed through comparison with typical ML models and sampling techniques. Meanwhile, the importance of input variables as well as the contribution score of input variables to the inherent decision-making process and specific predictive results were further explored. The competitive relationship between flexure and shear failures of CRCC was analyzed. Moreover, the influence of life-cycle corrosion behaviors on failure modes prediction of CRCC in different marine environments was also investigated. Analysis results indicate that the Matthews correlation coefficient , average f1 score, average precision, and accuracy of the EGA model were improved by about 5 % and 2–15 %, 4 % and 3–13 %, 5 % and 6–15 %, and 3 % and 2–9 % compared to the imbalanced dataset and typical sampling techniques, respectively. The average balanced accuracy, Matthews correlation coefficient , average f1 score, and accuracy of the EGA model were improved by about 9–29 %, 15–41 %, 14–28 %, and 9–26 %, respectively, compared with typical ML models.
Source: Engineering Structures
Authors: Hao Cheng, Weiwei Xie, Bo Yu
Institution: Guangxi University
Language: English
Subfield: Civil and Structural Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $4,620
OpenAlex ID: W4409610030