Forensic-based investigation-optimized extreme gradient boosting system for predicting compressive strength of ready-mixed concrete. Issue 1 (15th December 2022)

• Record Type: Journal Article
• Authors: Chou, Jui-Sheng; Chen, Li-Ying; Liu, Chi-Yun
• Is Part Of: Journal of computational design and engineering. Volume 10:Issue 1(2023)
• Page Start: 425
• Page End: 445
• Publication Date: 2022-12-15
• Subjects: metaheuristic optimization -- forensic-based investigation -- XGBoost -- ready-mixed concrete -- compressive strength -- prediction; Engineering -- Data processing -- Periodicals; Computer-aided design -- Periodicals; Computer-aided design; Engineering -- Data processing; Electronic journals; Electronic journals; Periodicals; 620.0042
• DOI: 10.1093/jcde/qwac133
• Languages: English
• ISSNs: 2288-4300

Abstract

Abstract: Regulations mandate testing concrete's compressive strength after the concrete has cured for 28 days. In the ideal situation, cured strength equals the target strength. Advanced estimation of concrete's compressive strength can facilitate quality management, improve safety, and present economic advantages in sustainable use. Basic statistical methods cannot effectively predict concrete's strength or its non-linear relationships with the proportions of its constituent materials. In this study, a baseline model for predicting concrete's compressive strength was constructed using a state-of-the-art machine-learning method. Most related studies have used sets of concrete mix design results concerning concrete specimens for laboratory-produced concrete specimens as training sets and have obtained simple models through regression; however, these models have been unsuitable for onsite prediction of the compressive strength of concrete with the various mix designs. Control over mix proportions is high in laboratories, resulting in low variation; onsite manual operation and environmental factors cause significant variations in assessment data. In this study, machine-learning techniques and a newly developed metaheuristic optimization algorithm were applied to big long-term data from 75 concrete plants to construct the optimal machine-learning model. Our self-developed forensic-based investigation algorithm was employed to fine-tune the hyperparameters of the extreme gradient boosting model and to improve the model's generalizability. The lowest mean absolute percentage error (MAPE) obtained using this model was 9.29%, which was smaller than the lowest MAPE achieved using the conventional simple regression with the water-to-binder (W/B) ratio (12.73%). The traditional method tends to overestimate the actual compressive strength. Finally, a convenient expert system was developed that facilitates the use of the proposed model by onsite engineers for quality management. This system expedites the judgment of whether a mixed design is reasonable, reducing production costs while maintaining the safety of concrete structures. It can be widely applied in practice and function as an effective decision-making tool. Graphical Abstract: