University of Transport Technology, Hanoi, Vietnam.
PLoS One. 2024 Nov 27;19(11):e0312531. doi: 10.1371/journal.pone.0312531. eCollection 2024.
Structurally, the lateral load-bearing capacity mainly depends on reinforced concrete (RC) walls. Determination of flexural strength and shear strength is mandatory when designing reinforced concrete walls. Typically, these strengths are determined through theoretical formulas and verified experimentally. However, theoretical formulas often have large errors and testing is costly and time-consuming. Therefore, this study exploits machine learning techniques, specifically the hybrid XGBoost model combined with optimization algorithms, to predict the shear strength of RC walls based on model training from available experimental results. The study used the largest database of RC walls to date, consisting of 1057 samples with various cross-sectional shapes. Bayesian optimization (BO) algorithms, including BO-Gaussian Process, BO-Random Forest, and Random Search methods, were used to refine the XGBoost model architecture. The results show that Gaussian Process emerged as the most efficient solution compared to other optimization algorithms, providing the lowest Mean Square Error and achieving a prediction R2 of 0.998 for the training set, 0.972 for the validation set and 0.984 for the test set, while BO-Random Forest and Random Search performed as well on the training and test sets as Gaussian Process but significantly worse on the validation set, specifically R2 on the validation set of BO-Random Forest and Random Search were 0.970 and 0.969 respectively over the entire dataset including all cross-sectional shapes of the RC wall. SHAP (Shapley Additive Explanations) technique was used to clarify the predictive ability of the model and the importance of input variables. Furthermore, the performance of the model was validated through comparative analysis with benchmark models and current standards. Notably, the coefficient of variation (COV %) of the XGBoost model is 13.27%, while traditional models often have COV % exceeding 50%.
从结构上看，侧向承载能力主要取决于钢筋混凝土（RC）墙。在设计钢筋混凝土墙时，必须确定抗弯强度和抗剪强度。通常，这些强度是通过理论公式确定的，并通过实验验证。然而，理论公式往往存在较大误差，且测试成本高、耗时。因此，本研究利用机器学习技术，特别是结合优化算法的混合 XGBoost 模型，根据现有实验结果进行模型训练，预测 RC 墙的抗剪强度。该研究使用了迄今为止最大的 RC 墙数据库，其中包含 1057 个具有各种横截面形状的样本。贝叶斯优化（BO）算法，包括 BO-高斯过程、BO-随机森林和随机搜索方法，用于改进 XGBoost 模型架构。结果表明，与其他优化算法相比，高斯过程是最有效的解决方案，它提供了最低的均方误差，并在训练集上实现了 0.998 的预测 R2、在验证集上实现了 0.972 的预测 R2、在测试集上实现了 0.984 的预测 R2，而 BO-随机森林和随机搜索在训练集和测试集上的表现与高斯过程一样好，但在验证集上的表现明显更差，具体来说，BO-随机森林和随机搜索在验证集上的 R2 分别为 0.970 和 0.969，而整个数据集包括 RC 墙的所有横截面形状。SHAP（Shapley Additive Explanations）技术用于阐明模型的预测能力和输入变量的重要性。此外，还通过与基准模型和现行标准的对比分析来验证模型的性能。值得注意的是，XGBoost 模型的变异系数（COV%）为 13.27%，而传统模型的 COV%通常超过 50%。
