【摘要】 基于试验或数值模拟的单一墩柱力学性能分析方法难以兼顾计算精度和效率，纯数据驱动的分析方法存在可解释性差和对数据依赖性强等问题。为此，本文通过研究钢筋混凝土（RC）墩柱力学性能试验数据、经验知识和机器学习的融合机制，提出了经验知识监督的RC墩柱力学性能神经网络（knowledge-guided neural network,KGNN）分析方法。首先，建立了包含761组RC墩柱拟静力试验样本的数据库；随后，基于经验知识分析了RC墩柱主要特征对其力学性能的影响规律，构建了相应的数学表征方法；最后，将RC墩柱试验数据及经验知识融入人工神经网络架构和训练过程，建立了高精度、可解释、可通用且不依赖大量训练数据的RC墩柱力学性能KGNN分析模型。本文提出的KGNN分析方法与纯数据驱动神经网络（BPNN）的结果对比表明：BPNN在测试集上表现更好，在分析墩柱承载力时均方根误差（E）和拟合系数（R2）分别为0.070和0.978,KGNN模型的E和R2分别为0.108和0.942；但由于BPNN所预测的墩柱特征对承载力的影响规律与经验知识并不吻合，即未能准确反映墩柱特征与其力学性能间的关系，BPNN模型发生了过拟合；而KGNN方法不仅可以快速准确获得RC墩柱力学性能，且预测规律与经验知识吻合较好，具有更高的可靠性和实用性。因此，融合试验数据与经验知识的神经网络有望成为一种新的RC结构力学性能分析方法。更多还原

【Abstract】 The mechanical performance analysis of reinforced concrete(RC) columns using only experimental or numerical methods usually faces challenges in balancing computational accuracy and efficiency, while purely data-driven methods suffered from poor interpretability and over-dependence on available data samples. To address this issue, an empirical knowledge guided neural network(KGNN)-based RC column analysis by investigating the fusion mechanism of empirical knowledge, test data and machine learning methods. A test database is firstly built based on 761quasi-static test specimens. In succession, the influence rules of primary characteristics of RC columns on their mechanical properties are analyzed based on empirical knowledge to formulate mathematical representations. Finally, the test data and empirical knowledge were implemented into the artificial neural network to develop high performance, explainable, generalizable KGNN model with only minor training samples. The result comparisons of the proposed KGNN method and the pure data-driven neural network(BPNN) demonstrate that although the BPNN slightly over the KGNN in terms of the load-carrying capacity prediction accuracy, with mean square error and correlation coefficient of 0.070 and 0.978 comparing to 0.108 and 0.942 of the KGNN. However, the results of the BPNN are not consistent with the empirical knowledge and further causes overfitting problem since it fails to capture the relationship between the characteristics and mechanical properties of RC columns. Fortunately, the KGNN method can not only quickly and accurately provide the mechanical properties of RC columns, but also present a higher consistency with the empirical knowledge with greater reliability and practicality. Through this work, the neural network-based methods integrating experimental data and empirical knowledge are expected to provide a novel analysis approach for RC structures.更多还原