关键词: 混凝土, 侵彻深度预测, 无偏特征选择, 贝叶斯优化

Abstract: Aiming at addressing the challenges of insufficient data, complex feature selection, and limited applicability of traditional empirical formulas in predicting concrete penetration depth, this paper proposes a prediction model of concrete penetration depth that leverages deep learning and optimization strategies. Initially, 218 sets of concrete penetration test data were collected, and through the integration of multi-stage empirical formulas, the dataset was expanded both in size and diversity. To enhance the accuracy of feature selection, an XGBoost (eXtreme Gradient Boosting) model combined with SHAP (SHapley Additive exPlanations) values was employed for feature screening and dimensionality reduction, thereby identifying key features that are correlated with penetration depth. Based on these selected key features, a multi-layer perceptron (MLP) model was constructed for penetration depth prediction. Further, Bayesian optimization was applied to fine-tune the hyperparameters of the MLP model, and K-fold cross-validation was used to evaluate its generalization capability. The results demonstrate that the proposed model performs better than the traditional empirical formula under various conditions, its prediction accuracy and stability are significantly improved, and it shows good generalization ability and adaptability. In addition, compared with the traditional empirical formulas, the model in this paper also shows better prediction effect in the applicable range of the empirical formulas.

Key words: concrete, penetration depth prediction, unbiased feature selection, Bayesian optimization