【作者】 赵启嘉；

【导师】 刘国彬；

【作者基本信息】 同济大学 ， 岩土工程， 2008， 硕士

【摘要】 近年来,随着上海城市轨道交通迅速发展,地铁建设工程的要求和难度越来越高,地铁深基坑受到地质条件、施工条件和外界其他因素的影响,成为了一个灰色、模糊、随机的系统。特别是在上海软土地区,地铁深基坑工程不仅涉及土力学中的基本强度、稳定、变形理论,还包含了土与结构的相互作用、渗流等水力学范畴的问题,此外,开挖时还会产生时空效应,所以这些因素给深基坑工程安全施工带来了不小的难度,而基坑变形控制作为控制基坑安全的一个重要措施,因此,准确而合理的预测基坑下一阶段的变形,对现场施工有着很好的实际意思。由于基坑工程动态非线性特点,加之内部和外部参数的不确定性,致使为基坑工程系统选择合适的建模预测方式是一个十分困难的事。因此,人们不得不寻求解决问题的新途径。神经网络由于具有自适应性、非线性和容错性强等特点,特别适合于处理各种非线性问题。它可以通过大量样本的学习来抽取出隐含在样本中的因果关系。由于深基坑工程具有高度非线性动态特性,各类监测数据间的联系也较为复杂,难以用具体的数学公式表示出,因此神经网络理论为深基坑预测工作提供了一条十分有效的途径。不同于数学建模的是,神经网络不需要建立数学模型,而是直接通过现场获得的监测数据来建立模型,从而避开了复杂的数学分析过程;同时,神经网络可以处理含有噪声和许多不确定因素的数据,以建立高度非线性的函数关系,事先不需要假设输出变量与输入变量之间的关系,而是通过样本的学习,实现输入与输出的非线性映射。本文选择动态递归神经网络来建立预测模型,同以前常用的静态网络(例如BP、RBF等)相比,更贴近深基坑动态非线性系统的特点,预测精度和效果明显优于静态网络。此外,由于基坑开挖会使得监测数据出现突变值,网络必须重新训练以适应新的样本数据,因此,本文还引入了混沌序列理论中的相空间重构技术,对动态递归神经网络预测模型的短期预测进行优化修正,建立新的预测模型,提出了基坑工程预测确定嵌入滞时τ和嵌入维数m二个重要参数的方法,并经工程实例分析,短期预测误差明显减小,整个新模型对基坑的短期预测得到了较好的效果。本文以上海地铁车站深基坑工程为背景,主要研究了以下内容:(1)从工程应用的角度,结合动态递归神经网络的特点,探讨了基坑工程变形预测建模的神经网络方法,阐述了Elman网络的基本原理、与静态网络相比的优势及特点,重点讨论了Elman动态递归神经网络的结构设计、数据的准备、训练网络及评价网络预测性能等内容;(2)研究神经网络预测方法在基坑变形上的适用性,总结了影响预测效果的几个因素,特别是数据样本的采集和数据变形“突变”的因素对模型预测效果的评价,并基于这些问题,引入了相空间重构的思想;(3)结合相空间重构理论,在Elman网络模型的基础之上,对时序数列进行重新定义,并结合工程实际,提出了相空间重构理论中的嵌入滞时τ和嵌入维数m二个重要参数,将修正后的预测结果与之前的预测结果进行对比,分析了预测对比效果;更多还原

【Abstract】 With the amelioration of urban rail transit network in Shanghai, projects of the metro construction have showed six trends. That is the foundation pits become deeper, the scale of constructions is getting larger, the distance between buildings and subway is getting closer, the time limit for constructions is more urgent, the geologic feature is more complex, the hidden troubles are more and more. How to accomplish the node target of completing 400km basic network in both ’quick and satisfactory’ way under the situation of large scale, leaping over style, high integration of risk becomes a question for the subway constructors in Shanghai. This paper concentrates on the evaluation of safety and deformation prediction of the foundation pits of subway stations in Shanghai soft soil from the needs of real projects.Neural network method is one of the most effective methods of deep foundation pit deformation prediction, which is an intelligent monitor method combined deformation forecast with control. The advantage of neural network lies in it provide a mathematics tool which can study and forecast by itself. In this paper, the Elman neural network is used to intelligently monitor deep foundation excavation. The Elman neural network is a typically dynamic recurrent neural network which is able to learn temporal patterns as well as spatial patterns. Therefore, the trained Elman neural network has the characteristics of the nonlinear and dynamic. At the same time Elman neural network avoids the drawback of traditional neural network which can not change the model structure real time and can not adapt to the abrupt change.The main ideas of the paper are listed as follows:(1) Design monitor system scientifically and put rigorous and effective monitor in practice so as well and truly show all kinds of pulse of structure and environment, and accordingly provide reliably and roundly basic information for design and construction;(2) Based on intelligent monitor and by virtue of the results obtained in this paper, we proposed the basic theory of choosing Elman neural network to model in the deep foundation pit intelligent monitor. Using Elman neural network overcome the disadvantage of slow convergence, easy to fall in to local infinitesimal and not having nonlinear dynamic characteristic when using the Bp neural network and RBF neural network which are largely used in the deep foundation pit deform prediction by far.(3) Change the method of non-continuous and static structure mechanics computation used in the past and make full use of artificial neural network predictive method to dynamically analysis every condition of deep foundation pit construction by time and space sequence and farther obtain the dynamic reliability;(4) Take the practical application into account, in this paper the mulit-data prediction method and time serial prediction method are used to intelligent monitor.更多还原