基于多机器学习的强夯有效加固深度预测对比研究

Comparative Study on Effective Reinforcement Depth Prediction of Heavy Tamping Based on Multiple Machine Learning

采用当前应用较为广泛的6种算法模型,即BP神经网络模型、SVM模型、RF模型、MLP模型、XGBoost模型以及朴素贝叶斯模型,对不同土质、不同夯击能量、不同夯锤面积下的强夯有效加固深度进行预测,采用均方根误差、平均绝对百分比误差、决定系数以及平均绝对误差4项指标对比各模型的预测准确率、误差及泛化能力,并结合工程实例数据做检验。结果表明,机器学习方法具有较高的预测精度和泛化能力,可通过少量且简单的强夯数据进行预测强夯有效的加固深度,能为强夯工程设计提供参考依据;在小样本的情况下,RF模型性能较为优异,其次为XGBoost模型与朴素贝叶斯模型,BP神经网络模型、MLP模型与SVM模型表现较差,建议在样本数据有限的情况下,优先选用RF模型进行强夯有效加固深度预测。

Six algorithm models,including BP neural network model,SVM model,RF model,MLP model,XGBoost model,and plain Bayesian model,are used to predict the effective reinforcement depth of rammers with different soil qualities,different tamping energies and different ramming areas,and four indicators,including root-mean-square error,average absolute percentage error,coefficient of determination and average absolute error,are used to compare the prediction accuracy,error and generalization ability of each model.These models are tested and examined with the data of engineering examples.The results show that,(a)the machine learning method has high prediction accuracy and generalization ability,and can predict the effective reinforcement depth of heavy tamping through a small amount of and simple heavy tamping data,which can provide a reference basis for the design of heavy tamping projects;and(b)in the case of small samples,the RF model performs better,followed by the XGBoost model and the plain Bayesian model,and the BP neural network model,the MLP model and the SVM model perform worse,so it is recommended to prioritize RF model for effective reinforcement depth prediction in case of limited sample data.