【作者】 邹承明；

【导师】 钟珞；

【作者基本信息】 武汉理工大学 ， 结构工程， 2003， 博士

【摘要】 列车在运行时产生的振动会通过周围地层向外传播，并进一步诱发建筑物的二次振动，对建筑物以及其中居民的工作和日常生活产生了很大的影响。如何分析振动对环境造成的影响已经成为一个很重要的课题。 轨道系统由铁轨、枕木和碎石垫层组成，列车在运行时车轮与轨道系统会产生相互作用，由于轮重的作用使轨道产生弯曲变形，弯曲变形转化成作用在地基上的荷载引起地基振动，这一振动通过周围地层向外传播，导致周围地面的振动。 为了得到神经网络的训练样本，除了利用现有的实测数据外，本文还考虑利用某些解析方法计算得到的数据作为样本。为此本文基于振动产生及传播的机理，利用了分层法的轨道交通引起地面振动的解析算法，通过解三维弹性体的波动方程得到波数—频域的解，离散化波数域得到地基土的动力响应特性，再通过快速Fourier逆变换得到空间—时间域的解。 为了利用神经网络进行有限元计算，本文给出了列车运行时移动荷载在时间域和空间域的分布表示，通过三维有限元方法的分析，详细推导出了轨道交通荷载引起地面振动的有限元计算中的刚度矩阵和质量矩阵，同时将动力有限元的计算问题通过矩阵变换，转换成了一个二次型的优化问题。 本文融合神经网络技术和遗传算法，对轨道交通引起的地面振动进行了正分析和反演，并利用计算机技术实现了地面振动的三维动画模拟，这主要包括： (1)研究了压缩映射遗传算法，利用Banach空间的不动点定理证明了压缩映射遗传算法可以收敛到全局最优值，弥补了标准遗传算法不能收敛至全局最优值的不足。另外还详细论述了该算法中的编码、选择、交叉和变异等遗传操作。为后续与神经网络融合应用作理论准备。 (2)研究了BP神经网络的结构、训练算法，并将该算法应用到土的本构关系的建模之中。考虑到BP神经网络训练时间长、容易收敛到局部最优的不足，本文将压缩映射遗传算法应用到BP神经网络的训练之中，使其较快地收敛到全局最优值。 (3)研究了Hopfield反馈神经网络的结构和稳定性，并深入讨论了稳定性分析的Lyapunov第二方法(亦称为直接法)，利用Lyapunov方法分析了Hopfield神经网络改进网络—TH神经网络的稳定性。本文详细论述了将TH神经网络应用到二次型优化问题上的方法，并首次将该方法用来计算前面得武汉理工大学博士学位论文到的动力有限元计算的二次型。在仿真计算中，首次将压缩映射遗传算法应用到了搜索TH神经网络的平衡点(二次型的最优值)上。 (4)研究了自递归神经网络的结构及训练算法，利用LyaPunov方法研究了自递归神经网络训练算法的收敛性和稳定性，并首次利用自递归神经网络来反演轨道交通引起地面振动的振动场。考虑到自递归神经网络的结构影响到其反演效果，本文利用压缩映射遗传算法来搜索其最佳的结构。 (5)论述了轨道交通引起地面振动智能分析软件中运用的面向对象方法和OpenGL技术，研究了OpenGL的三维体绘制算法，最后实现了轨道交通引起地面振动的振动场三维模拟。 本文最后总结了研究结论并展望进一步的工作，认为包括模糊逻辑、灰色理论等多种智能方法和神经网络、遗传算法的融合应用到轨道交通引起地面振动的分析之中是一种新的可行方法，三维实时动画模拟也是研究方向。更多还原

【Abstract】 The train-induced vibration will spread outward through the medium of around strata, and induce the secondary vibration of construction in further, which brings great influence to construction and people’s daily life. How to analysis the environment effect cased by vibration has become a very important problem.The track system is composed of rail-, sleeper ballast and substrate. When the train moves, the wheel and’the track system will interact each other. Because of the action of the wheel, the track may bright bent transformation, and the transformation change into load, which act on the foundation. Then the ground will be vibration caused by the transformation spreads in the ground.In order to get the neural network’ training samples, the paper will use the data computed by some analytic method as training samples besides using existing data which measured by in locale. The paper based on the mechanism of the vibration’s production and transmission, uses the analytic method of finite elastic layer to calculate the transmission of the ground vibration caused by track traffics. Then gets the result of wave numbers-frequency field through solving the wave equation of the three-dimension elastic layer. Then gets the dynamic response characteristic of the ground through dispersing the wave numbers. Finally gets the result of the space-time field through using reverse fast Fourier transform.The paper gives the expressions of the moving load in the time-space field. The paper gives stiffness matrix and mass matrix in finite element analysis of the ground vibration problem caused by track traffic by using three-dimension finite element method. Because three-dimension finite element analysis is very difficult, it is transmitted into a quadratic programming, which can be calculated by neural network.The paper analyzes the ground vibration caused by track traffic in two ways by neural network and genetic algorithm fusion. One side is forward analysis, on the other side is inversion. And the paper implements the three-dimensional dynamic simulation of the ground vibration. It mainly includes five parts.(1) The paper has studied the contractive mapping genetic algorithm, which can be converged to a global optimal solution proved by fixed-point theorem of Banach space. It remedies the shortage of standard genetic algorithm. Furthermore, the paper discusses the genetic operation in the algorithm such as the coding, select, cross over and mutation etc. It will be help to the next work about fusion with the neural network.(2) The paper has studied the structure and training algorithm of BP neural network. And it is applied to the modeling of constitutive model. Considering the shortage of BP neural network such as too long training time and easily converging the local optimal value, the paper applies the contractive mapping genetic algorithm to BP neural network1 training, which make it consequently converges to the globe optimal value more quickly.(3) The paper has studied the structure and stability of Hopfield feedback neural network. And it uses the Lyapunov method to analysis the stability of the TH neural network, which id mended from Hopfield neural network. The paper applies the TH neural network to a quadratic programming, which is transmitted from finite element computing. Finally the paper applies the contractive mapping genetic algorithm to search the equilibrium point of TH neural network in a quadratic programming computing for the first time.(4) The paper has studied the structure and training algorithm of self-recursion neural network. It also studies the convergence and stability of self-recursion neural network’ training algorithm by using Lyapunov method. Then the paper use inversion method to simulate the vibration field of ground vibration caused by track traffic by self-recursion neural network for the first time. Considering self-recursion’ structure can work on its inversion effect, so the paper uses contractive mapping genetic algorithm to search its optimal struct更多还原