打入桩轴向极限承载力预测的机器学习方法

Prediction of Ultimate Axial Load-carrying Capacity for Driven Piles Using Machine Learning Methods

在复杂的工程地质环境中,打入桩的轴向承载力的预测是设计和施工的一个重大挑战。本研究将利用机器学习工具一极端梯度提升算法(XGBoost)预测打入桩的轴向承载力,对桩身最大压缩应力(MCS)、最大拉伸应力(MTS)和每英尺锤击数(BPF)进行了研究,并与反向传播神经网络(BPNN)和随机森林(RF)算法进行了比较分析。利用美国北卡罗来纳州的桩数据库中4000多个数据集生成训练和预测样本。利用均方根误差(RMSE)、平均绝对误差(MAE)和线性相关系数(R^2)等性能指标对三种回归算法的有效性进行了验证。结果表明,与其他两种方法相比.XGBoost算法在解决桩、锤、土以及施工工艺等复杂非线性问题时,具有更高的稳定性和预测精度。XGBoost算法作为一种可靠的机器学习算法,可以为其他类似桩基工程轴向承载力的预测提供参考。

Prediction of axial load-carrying capacity for driven piles is one of the challenging tasks in the design and construction, especially it is utilized in complex geological environment.In this research, a promising practical machine learning tool known as extreme Gradient Boosting (XGBoost) model is applied for prediction of pile bearing capacity.Maximum Compressive Stress (MCS), Maximum Tensile Stress (MTS) and Blow Per Foot (BPF) are the key targets.Moreover, the other two methods: Back-Propagation Neural Network (BPNN) and Random Forest (RF) are developed for comparison purposes.A pile database more than 4 000 data sets are used into generate training and testing samples.The validation and comparison of the three regression models are evaluated by several performance indices-the Root Mean Square Error (RMSE), the Mean Absolute Error (MAE) and the coefficient of determination (R^2).The results show that XGBoost algorithm has higher prediction accuracy and stability than the rest two methods for solving complicated and nonlinear problems among pile engineering, hammer, soil, construction technology and pile drivability.It is concluded that XGBoost as a reliable and accurate technique could potentially be further employed for target variables estimation and provide significant references for other similar pile engineering.