基于BP神经网络的参数反演在基坑变形预测中的应用  

Application of soil parameter inversion method based on BP neural network in foundation pit deformation prediction

作　　者：马豪豪[1] 袁帅 张志增[1] 田亚辉 董森森 Ma Hao-Hao;Yuan Shuai;Zhang Zhi-zeng;Tian Ya-hui;Dong Sen-Sen(Department of Architecture and Engineering,Zhongyuan University of Technology,Zhengzhou 450007,China;Department of Geotechnical,Highway School,Chang’an University,Xi’An,710054,China;China Railway Construction Engineering Fifth Construction Co.,Ltd,Guangzhou,510000,China;不详)

机构地区：[1]中原工学院建筑工程学院,郑州451100 [2]长安大学公路学院,西安710054 [3]中铁建工集团第五建设有限公司,广州510000 [4]安岩智能科技(常州)有限公司,常州213000

出　　处：《Applied Geophysics》2023年第3期299-309,349,共12页应用地球物理（英文版）

基　　金：The National Natural Science Foundation of China(Grant number 52104157);Natural Science Foundation of Henan Province(Grant number:222300420596);NSFC-Henan Province Talent Training Joint Fund(Grant number:U1204509).

摘　　要：当基坑变形过大时,为了能及时采取措施以确保基坑安全,一种能精确预测基坑变形的方法至关重要。常用的有限元模拟计算受土体参数影响较大,不能反映基坑变形的动态变化。本文以某深基坑工程为依托,基于实际土层参数采用正交试验设计了64组代表性土体参数组合,利用基坑的三维有限元模型计算了各参数组合的位移值,如支护结构的最大水平位移值和地表沉降值等。以土体参数和计算得到的变形值作为输入值和输出值训练设计的BP神经网络模型。模型训练完成以后,以实际监测得到的基坑变形数据为输出反演能够反映基坑动态变化的土体参数,然后利用有限元模型计算得到下一阶段的变形数据。在工程中的应用结果表明,通过BP神经网络模型反演得到的土体参数能更好的反映深基坑受力状态,得到的预测结果与监测值吻合较好,验证了该方法的准确性和可行性。When significant deformation occurs in a foundation pit,it is critical to have an accurate method for predicting this deformation.This is necessary for enacting timely safety measures.Unfortunately,finite element simulations,which are strongly affected by soil parameters,fail to reflect the dynamic deformation of foundation pits during excavation.To address this,we used the actual soil parameters of a deep foundation pit to design 64 representative combinations of soil parameters through orthogonal testing.Using a three-dimensional(3D)finite element model of the foundation pit,we obtained displacement values for each parameter combination.These included the maximum horizontal displacement of the support structure and the surface settlement value.Subsequently,we developed a backpropagation(BP)neural network model.We trained this model using the soil parameters of each combination as input and the deformation values obtained from the 3D finite element model as output.Once the model was trained,we inverted the soil parameters,reflecting the dynamic deformation of the foundation pit by using actual monitoring data.This process allowed us to obtain the deformation data for the next excavation stage.Results showed that the soil parameters obtained via the BP neural network model effectively reflected the stress state of the deep foundation pit.Moreover,the prediction of the foundation pit deformation aligned well with the monitoring data,which validates the accuracy and feasibility of our method.

关 键 词：深基坑 参数反演 有限元模拟 BP神经网络 变形预测 

分 类 号：TU753[建筑科学—建筑技术科学]