基于随机森林算法的砂土液化预测方法  

The method of predict sand liquefaction based on random forest algorithm

作　　者：彭刘亚[1] 解惠婷 冯伟栋[1] PENG Liu-Ya;XIE Hui-Ting;FENG Wei-Dong(Anhui Earthquake Engineering Institution,Anhui Earthquake Administration,Hefei 230031,China)

机构地区：[1]安徽省地震局安徽省地震工程研究院,安徽合肥230031

出　　处：《物探与化探》2020年第6期1429-1434,共6页Geophysical and Geochemical Exploration

基　　金：中国地震局三结合课题(3JH202002013)。

摘　　要：砂土液化的影响因素较多且复杂。以唐山大地震的72个场地的实测液化样本数据为例,在不丢失任何信息的前提下,选取了8个砂土液化的判别指标,通过计算样本数据的Gini系数,采用CART算法的决策树对数据的特征属性进行划分。在此基础之上,通过增加多个决策树构造随机森林的方式,在一定程度上降低了单个决策树学习过度造成的过拟合风险,同时,通过10轮交叉验证的方式确定了决策树的最大高度为5,随机森林中决策树的个数为20时,模型的效果达到最佳。研究结果表明,与抗震设计规范中的标贯试验法判别公式相比,决策树模型和随机森林模型的训练结果和预测结果有显著提高,尤其是随机森林模型在训练样本和预测样本上均没有出现误判,稳定性更高。Among a variety of complicated factors that are related to sand liquefaction,8 discriminant factors have been picked out of 72 samples in the earthquake event happened in Tangshan without losing any tiny but useful information.By calculating Gini coefficient with CART algorithm,a decision tree has been undertaken to divide the features of original sample dataset.Moreover,in order to reduce overfitting risk of a single decision tree,random forest with multiple trees have been created.Meanwhile,with 10-fold cross validation,best estimators with 5 max-depth and 20 trees can perform with much more stable and reliable results.The research shows that,compared to standard penetration test from Code for seismic design of buildings,both decision tree and random forest have a better predicting precision,especially there have been no false classifications with higher stability using random forest model.

关 键 词：砂土液化 判别指标 决策树 随机森林 

分 类 号：P631.4[天文地球—地质矿产勘探]