基于深度学习的水工混凝土空隙结构研究

Void Structure of Hydraulic Concrete Based on Deep Learning

摘要:

考虑定量化表征水工混凝土的三维空隙结构几何形态参数，系统分析空隙结构几何形态参数的分布规律及内在相关性。同时，采用SHAP法开展可解释性分析，揭示输出目标与输入特征之间的复杂映射关系；基于深度神经网络（DNN）构建水工混凝土空隙形态特征与总空隙率的定量关联模型。结果表明：等效半径与磨圆度、空隙数量与熵、表面积与磨圆度之间存在较强的正相关性，欧拉特征数与分形维数呈现负相关性；对特征重要性进行排序，输入特征压缩至5个；构建的DNN预测模型训练集和测试集的决定系数分别为0.90和0.93，预测精度较高；本研究建立的空隙结构几何形态参数表征体系为水工混凝土性能优化提供了量化分析的工具模型与理论支撑。

Abstract:

The geometric characteristics of the three-dimensional void structure in hydraulic concrete were quantitatively characterized， and the distribution patterns and internal correlations of void geometric parameters were systematically analyzed. In addition， interpretability analysis was performed using the Shapley additive explanations（SHAP） method to elucidate the complex mapping relationships between the output target and input features. A quantitative association model between void morphological features and total porosity was constructed using a deep neural network（DNN）. The results show that a strong positive correlation is observed between equivalent radius and roundness， number of voids and entropy， as well as surface area and roundness. In contrast， a negative correlation is found between Euler characteristic and fractal dimension. Feature importance is ranked， and input features were reduced to five. The constructed DNN prediction model achieves correlation coefficient of 0.90 and 0.93 for the training and testing sets， respectively， indicating high prediction accuracy. The geometric characterization system for void structures established in this study provides a quantitative analytical tool and theoretical support for the performance optimization of hydraulic concrete.