基于集成学习和LEM的折线型边坡稳定性评价方法  

Evaluation for Broken Line Slope Stability Based on Ensemble Learning and LEM

作　　者：邓子昊 张黎明 徐兴华[5] 吕庆 Deng Zihao;Zhang Liming;Xu Xinghua;LüQing(College of Civil Engineering and Architecture,Zhejiang University,Hangzhou 310058,China;Center for Balance Architecture,Zhejiang University,Hangzhou 310058,China;Architectural Design&Research Institute of Zhejiang University,Hangzhou 310012,China;Mining College,Guizhou University,Guiyang 550025,China;Zhejiang Institute of Geology and Mineral Resources,Hangzhou 310007,China)

机构地区：[1]浙江大学建筑工程学院,浙江杭州310058 [2]浙江大学平衡建筑研究中心,浙江杭州310058 [3]浙江大学建筑设计研究院有限公司,浙江杭州310012 [4]贵州大学矿业学院,贵州贵阳550025 [5]浙江省地质矿产研究所,浙江杭州310007

出　　处：《地球科学》2024年第11期4216-4224,共9页Earth Science

基　　金：浙江省重点研发计划(No.2021C03159);国家自然科学基金面上项目(No.42277132);浙江大学平衡建筑研究中心配套资金资助项目(No.K-20212721);浙江省基础公益研究计划(No.LGF21D020001);自然资源部浙江地质灾害野外观测研究站项目(No.ZJDZGCZ-2021).

摘　　要：传统边坡稳定性力学分析方法计算效率有限且往往需要借助专业软件,不利于推广.机器学习作为一种高效分析手段可应用于边坡稳定性评价.基于随机生成的大量折线型边坡样本,通过极限平衡法(LEM)求解安全系数,从而构建边坡安全系数数据库,通过集成神经网络模型建立LEM代理模型.分别采用Bagging和AdaBoost.R2两种算法构建集成神经网络,建立折线型边坡安全系数预测模型,通过实际边坡工程案例进行了验证,并与单神经网络进行了对比.通过ROC曲线分析法评价各个模型性能,确定合理的安全系数阈值.结果表明两种集成模型性能显著优于单神经网络模型,其中单神经网络模型的AUC值为0.826,AdaBoost.R2模型为0.893,Bagging模型为0.929,更能准确辨别边坡的稳定性情况.提出的方法能快速、准确地评价折线型边坡稳定性,为区域性大量边坡的稳定性快速评价提供工具.Conventional mechanical analysis methods for slope stability have limited computational efficiency and need professional software.Machine learning,as an efficient analysis method,can be applied to slope stability evaluation.Abundant broken line slope samples are randomly generated and the corresponding factors of safety are solved by the limit equilibrium method(LEM)in this paper,so as to build a slope factor of safety database,and the LEM surrogate model is established by integrating neural network models.Two ensemble algorithms,Bagging and AdaBoost.R2,are used to establish a neural network ensemble model to predict factor of safety,which is verified by practical slope engineering cases,contrasting with single neural network model.The performances are evaluated by ROC curve analysis method,and reasonable threshold of factor of safety is determined.Results show that two ensemble models are significantly better than the single neural network model.While the AUC value of the single neural network model is 0.826,the AdaBoost.R2 model is 0.893,and the Bagging model can recognize slope stability situation better with value of 0.929.The proposed method can evaluate broken line slope stability quickly and accurately,providing a tool for rapid stability evaluation of a large number of regional slopes.

关 键 词：集成神经网络 极限平衡法 折线型边坡 边坡稳定性 ROC曲线分析法 滑坡 工程地质. 

分 类 号：P694[天文地球—地质学]