【作者】 岳荣花；

【导师】 黄腾；

【作者基本信息】 河海大学 ， 大地测量学与测量工程， 2007， 硕士

【摘要】 沉降监测贯穿着建筑物设计期、施工期和运营期的整个过程，是不可忽视的问题，工程技术人员都给予极大的重视和严格的分析。沉降预测方法很多，主要分为两大类，一类是理论方法，一类是基于实测数据的实测数据分析法。由于影响沉降的因素很多且具有不确定性，而各个因素对沉降的影响均表现在沉降数据上，因此第二类方法得到了广泛的应用。本文针对传统沉降预测模型应用中存在的缺陷，研究了小波神经网络沉降预测模型，对基于传统BP算法的小波神经网络进行了优化并应用于工程实例中。主要研究内容如下：1．研究了小波神经网络算法。对基于传统BP算法的小波神经网络局限性进行分析，引入了一种新的无约束优化线搜索，并推导出新线搜索下的DY-HS杂交共轭梯度反向传播算法来训练小波神经网络，改善了基于BP算法小波神经网络易产生收敛于局部极小及速度慢的问题。2．探讨了小波神经网络初始权值的自相关修正法。即将小波神经网络的初始参数设置和小波类型、小波时频参数和学习样本等联系起来的小波神经网络的初始参数设置方法。该方法有别于传统网络初始权值的随机赋值，提高了网络的稳定性及收敛精度。3．优化了小波神经网络结构。将新线搜索下的DY-HS杂交共轭梯度反向传播算法和参数初始值的自相关修正法结合运用，采用一种简单的变结构方式来调节隐含层节点数，推导出改进小波神经网络。改进的小波神经网络是对传统基于BP算法的小波神经网络的优化，有效克服了基于BP算法的小波神经网络存在的一些缺陷。4．利用BP神经网络、基于BP算法的小波神经网络和改进的小波神经网络建立沉降时间序列预测模型。将训练样本的选取分为累积沉降和间隔沉降两种方案，对不同工程实例进行预测。通过对比分析预测结果，改进的小波神经网络模型优于其它两种网络模型，且根据不同的沉降状态选取不同训练方案可以获取更高的预测精度。更多还原

【Abstract】 Settlement observation involved all through the design period,construction period, operation period of engineering structure. All engineering and technical staffs have paid much attention and strict analysis to the important problem which should not be neglected. Settlement forecast methods could be divided into theoretical methods and experimental method that based on calculating field data. There are many influence factors of settlement which are uncertain. Because of all the factors have influenced settlement in subsidence observation, the second methods have been widely used. This paper discussed the defects of traditional settlement prediction models, researched wavelet neural network of settlement prediction model, then put forward measures and approaches for improved wavelet neural network based on BP algorithm and applied in engineering examples. The main contents of this paper are as follows:1. The wavelet neural network arithmetic is studied. Uniting the study of artificial neural networks and its combination with wavelet transform and BP networks’ fruit before, the defects of wavelet neural network based on BP algorithm were analyzed, and a new unconstrained line search is introduced to improve conjugate gradient methods, then derived DY-HS crossing conjugate gradient back propagation algorithm under the new line search to train wavelet neural network, which solve the problem of easily relapsing into local minimization of classical wavelet neural network based on BP algorithm.2. The self-correlation correction for initial weights of wavelet neural network is discussed. This method integrates the setting of initial parameters with the wavelet type, time-frequency parameters of the wavelet and the training samples, and it is different from traditional method which get initial weights randomly and the stability and convergence precision of network is improved.3. The configuration of wavelet neural network is optimized. The improved waveletneural network based on DY-HS crossing conjugate gradient back peopagation algorithm is put forward with colligating improved method above which is studied. And a simple form of varied structure is introduced to regulate the node of hidden layer of network. The improved wavelet neural network optimize the classical wavelet network based on BP algorithm, can effectively conquer the shortages of classical wavelet network based on BP algorithm.4. The BP network,BP wavelet network and improved wavelet network is used insetting up the settlement forecasting model. The selection of training samples were adopted into two schemes: the total settlement and the interval settlement, then applied the two schemes in forecasting different engineering examples. Compared with various forecasting results, the improved wavelet network model was better than other model, selecting training different samples according to different sinking periods would acquired higher prediction precision.更多还原