基于符号回归的受弯破坏矩形RC剪力墙变形能力预测

Prediction of deformation capacity of rectangular RC shear walls subjected to bending damage based on symbolic regression

为克服传统回归方式的局限性,同时解决机器学习模型可解释性差的问题,利用SHAP法与符号回归建立了具有更高精度的经验公式。为此,基于已有的试验数据库,建立了一个包含119个弯曲破坏矩形RC剪力墙的数据库。利用多种机器学习算法建立回归模型来预测剪力墙极限位移,其中XGBoost模型回归效果最佳。利用SHAP法对模型进行了分析与解释,并基于SHAP值对特征参数进行了筛选,以提升符号回归的效率与预测性能。提出的回归方法有助于克服传统回归方式的缺陷,与已有经验公式的对比结果表明,提出的经验公式具有更高的预测精度。此外,对符号回归模型所选用的特征参数进行了参数分析,并用XGBoost模型结果进行对比,以得到各特征参数在不同模型间的差异。结果表明,符号回归模型与XGBoost模型中各特征参数对极限位移的影响趋势基本一致,且符号回归模型具有更好的泛化性能与可解释性。

To overcome the limitation of traditional regression approach and address the poor interpretability of machine learning models,this paper uses the SHAP method and symbolic regression to establish an empirical formulation with a higher accuracy.For this purpose,a database containing 119 flexural damaged rectangular reinforced concrete(RC)shear walls was established based on the existing literature.Multiple machine learning algorithms were used to establish regression models to predict the ultimate displacement of shear walls,among which the XGBoost model had the best regression effect.The model was analyzed and interpreted using the SHAP method,and the feature parameters were filtered based on the SHAP values to improve the efficiency and prediction performance of the symbolic regression.The regression method proposed in this paper overcame the defects of the traditional regression approach,and the comparison results with the existing empirical formulas showed that the empirical formula proposed in this paper had a higher prediction accuracy.In addition,a parametric analysis of the feature parameters selected for the symbolic regression(SR)model was conducted and compared with the results of XGBoost model to obtain the differences of each feature parameter among different models.The results showed that the trend of the influence of each feature parameter on the limit displacement in the SR model and the XGBoost model was basically the same,and the SR model had better generalization performance and interpretability.