基于ML预测不确定性的边坡失稳时空概率评估方法与应用研究  

Spatial-temporal probability assessment method for slope instability based on ML prediction of uncertainty and its application

作　　者：邓志兴 谢康 李泰灃 苏谦[3] 肖宪普 王业顺[1] 刘宝 Deng Zhixing;Xie Kang;Li Taifeng;Su Qian;Xiao Xianpu;Wang Yeshun;Liu Bao(Central South University,Changsha 410075,China;Railway Engineering Research Institute of China Academy of Railway Sciences Co.,Ltd.,Beijing 100081,China;Southwest Jiaotong University,Chengdu 610000,China;Shijiazhuang University of Railways,Shijiazhuang 050000,China;China Railway Design Corporation,Tianjin 300308,China)

机构地区：[1]中南大学,湖南长沙410075 [2]中国铁道科学研究院集团有限公司铁道建筑研究所,北京100081 [3]西南交通大学,四川成都610000 [4]石家庄铁道大学,河北石家庄050000 [5]中国铁路设计集团有限公司,天津300308

出　　处：《土木工程学报》2024年第9期109-122,共14页China Civil Engineering Journal

基　　金：时速400公里高速铁路基础设施适应性关键技术研究(P2021G053)。

摘　　要：为解决机器学习(Machine Learning,ML)边坡位移预测中的不确定性及预测信息挖掘深度不够等问题,提出基于ML预测不确定性的边坡失稳时空概率评估方法。首先,基于Bootstrap算法度量ML边坡位移预测中的不确定性,利用GRU算法预测边坡单一点位的位移时间特征、Kriging算法插值边坡位移空间分布,建立Bootstrap-GRU-Kriging边坡位移时空不确定性预测模型;接着,利用可靠度理论挖掘位移时空不确定性预测结果,建立边坡失稳时空概率评估模型,基于最不利原则提出边坡全断面位移-失稳概率二元耦合分析指标DP;最后,依托杭绍台铁路硅藻土边坡试验段验证模型的有效性。结果表明:边坡位移时空不确定性预测模型兼具ML边坡位移预测不确定性量化和时空分布预测功能,边坡失稳时空概率评估模型不仅能得到边坡失稳概率空间分布及时变特性,还能实现边坡整体安全的高效评价;所提方法构造的预测区间清晰可靠,各测点预测区间覆盖率PICP均值高于96%,DP指标可响应边坡变形过程中外界所造成的影响。研究成果可为边坡位移预测及超前评估理论研究提供参考。To solve the problems of uncertainty and insufficient depth of prediction information exploration associated with Machine Learning(ML)for slope displacement prediction,a spatial-temporal probability assessment method of slope instability based on ML prediction of uncertainty is proposed.Firstly,the Bootstrap algorithm is utilized to measure the uncertainty in ML for slope displacement prediction,the GRU algorithm is used to predict the displacement-time characteristics of a single point of the slope and the Kriging algorithm is employed to interpolate the displacement spatial distribution,in order to establish the Bootstrap-GRU-Kriging prediction model for the spatial-temporal uncertainty of slope displacement;then,based on the spatial-temporal uncertainty of predicted displacement results exploited by the reliability theory,the spatial-temporal probability assessment model of slope instability is established,and based on the most unfavorable principle the binary coupling analysis index DP of displacement-instability probability of the whole slope section is proposed;finally,the validity of the model is verified by the test section of diatomite slope of Hang-Shao-Tai Railway.The results show that:the prediction model for spatial-temporal uncertainty of slope displacement possesses the function of ML for uncertainty quantification of slope displacement prediction and the function of spatial-temporal distribution prediction,and the spatial-temporal probability assessment model of slope instability can not only obtain the spatial distribution of slope instability probability and its time-varying characteristics,but also realize the efficient evaluation of overall slope safety;the proposed method provides clear and reliable prediction intervals,and the mean value of Prediction Interval Coverage Probability(PICP)for each measurement point is higher than 96%.The DP index can reflect the external influence in the process of slope deformation.The research results can provide a reference for theoretical research on slope

关 键 词：边坡位移 机器学习 不确定性度量 可靠度理论 时空概率评估 DP指标 

分 类 号：U213.1[交通运输工程—道路与铁道工程]