基于BP神经网络的UHPC直剪承载力预测模型

Prediction Model of Direct Shear Strength of UHPC Based on BP Neural Network

为提供准确的UHPC直剪承载力,以指导UHPC结构设计,建立一种基于BP神经网络的UHPC直剪承载力预测模型。该方法基于机器学习中的反向传播人工神经网络(BP-ANN),搜集现有相关试验数据并建立数据库,将混凝土抗压强度、受剪面积、纤维特征参数、钢筋参数和侧向约束应力指定为输入特征参数,将直剪承载力指定为输出量,利用数据库对BP-ANN模型进行训练。将模型预测值与试验实测值和现有计算模型的结果进行对比,并采用SHAP算法对各参数重要性进行分析。结果表明:BP-ANN模型具有更好的预测效果,其相关系数R2达到0.953,平均绝对误差MAE为1.015,模型训练结果理想,可应用于实际的数据处理分析;侧向约束应力对直剪承载力的影响最大,钢筋参数影响最小。

To provide accurate UHPC direct shear strengths for the design of UHPC structures,it is necessary to establish a UHPC direct shear strength prediction model with high versatility and accuracy.The prediction model is based on the back-propagation artificial neural network(BP-ANN)in machine learning,in which the existing relevant experimental data are collected to build a database,the data pertaining to concrete compressive strength,shear surface area,fiber characteristic parameters,reinforcement parameters and lateral restraint stress are specified as inputs,and the direct shear strengths the outputs.The database is employed to train the BP-ANN model and analyze the importance of the parameters by SHAP,aiming to figure out the influence of the parameters on the direct shear strengths.Compared with the existing UHPC bearing capacity calculation model,the BP-ANN model displays a better prediction effect,with a correlation coefficient(R 2)and an average absolute error(MAE)of 0.953 and 1.015,respectively,indicating that the model training results are ideal,which can be used for actual data processing and analysis.The lateral restraint stress exerts the greatest influence,while the reinforcement parameters have the least influence on the direct shear strength.