摘要
为了预测深基坑支护桩水平变形的长期发展规律,在卷积神经网络(convolutional neural network,简称CNN)数据空间特征提取基础上,结合长短时记忆神经网络(long and short term memory,简称LSTM)分析数据的时序性和注意力机制(attention mechanism,简称AM)的划分特征权重,构建了能够预测支护桩变形的AM-CNN-LSTM模型。以北京地区某深基坑工程为背景,基于灰色关联方法明确了影响支护桩最大变形的因素,通过构建的模型分析支护桩的单点变形规律,并与反向传播神经网络(back propagation neural network,简称BPNN)、CNN和传统CNN-LSTM模型的预测所得结果进行比较分析。研究结果表明:支护桩最大变形值与深基坑开挖深度、临空天数、支撑内力、土壤性质、桩的尺寸和嵌固深度等因素关联度较高;AM机制显著提升了初始数据信息挖掘深度和变形预测精度,通过梯度下降法不断更新直至满足误差要求;与BPNN、CNN及CNN-LSTM模型相比,AM-CNN-LSTM模型的应用对于支护桩的长期变形预测稳定性较好;通过与实测数据对比,AM-CNN-LSTM模型的预测精度误差在5%~10%以内。
In order to predict the long-term development pattern of horizontal deformation of deep foundation pit support piles,an AM-CNN-LSTM model capable of predicting the deformation of support piles was constructed based on spatial feature extraction of convolutional neural network(CNN)data combined with long and short term memory neural network(LSTM)to analyze the temporal nature of the data and the divided feature weights of attention mechanism(AM).In the context of a deep foundation pit project in Beijing,the factors affecting the maximum deformation of the supporting piles are clarified based on the gray correlation method.The constructed model was used to analyze the single-point deformation pattern of the supporting pile and to compare and analyze the results obtained from the predictions of back propagation neural network(BPNN),CNN and traditional CNN-LSTM models.The results show that the maximum deformation value of the supporting piles is highly correlated with the excavation depth of the deep foundation pit,the number of days of proximity,the internal force of the support,the nature of the soil,the size of the piles,and the embedment depth.The AM mechanism significantly improves the initial data information mining depth and deformation prediction accuracy,which is continuously updated by the gradient descent method until the error requirements are satisfied.Compared with BPNN,CNN and CNN-LSTM models,the application of AM-CNN-LSTM model is more stable for long-term deformation prediction of supporting piles.By comparing with the measured data,the prediction accuracy of the AM-CNN-LSTM model is within 5%to 10%error.
作者
李涛
舒佳军
王彦龙
陈前
LI Tao;SHU Jia-jun;WANG Yan-long;CHEN Qian(School of Mechanics and Civil Engineering,China University of Mining and Technology-Beijing,Beijing 100083,China)
出处
《岩土力学》
EI
CAS
CSCD
北大核心
2024年第S01期496-506,共11页
Rock and Soil Mechanics
基金
国家自然科学基金青年科学基金项目(No.51508556)
中央高校基本科研业务费专项资金项目(No.2022YJSLJ15)
越崎青年学者资助项目(No.800015z1166)
作者简介
第一作者:李涛,男,1981年生,博士,副教授,博士生导师,主要从事城市地下工程方面的教学和研究工作。E-mail:lit@cumtb.edu.cn
