【作者】 徐振东；

【导师】 李颖；

【作者基本信息】 吉林大学 ， 测试计量技术及仪器， 2004， 硕士

【摘要】 本文是在完成国家地质实验中心2002年基于GIS的多源地学信息整合处理技术算法设计子课题基础上完成的。主要研究人工神经网络数学模型和计算方法的计算机程序实现，为实现复杂地质信息的非线性整合处理提供技术支持。 人工神经网络(Artificial Neural Network，简称ANN)是一种大规模自适应非线性动力学系统，能实现非线性映射、模式识别、函数逼近、聚类分析、数据压缩、优化设计等功能，同时又具有很强的稳定性、收敛性、鲁棒性等良好性质，在各种信息处理领域有着广泛的应用。 地质工作中存在着大量的非线性问题，对地质勘探资料进行综合分析与分类，准确预测矿产资源的储量和分布等都涉及到多资料的整合处理，因此研制理想的非线性整合处理方法是非常必要的，而人工神经网络的一些良好特性，恰好能满足地质工作的需求，因此越来越多的人把研究工作的重点纷纷转向基于人工神经网络的非线性数学模型，以期更好地解决地学中的复杂问题。 目前，这方面的研究已有了部分成果，但大多只局限于应用一种固定的网络模型来解决某一具体的应用，还远远不能满足地质勘探的多种需求，本课题的研究目的就是要建立多种网络模型，并在算法实现过程中使其具有通用性，以满足各种应用的需要。 本文的具体工作首先是对人工神经网络的三种常用模型—Bp网、Hopfield网和Kohonen网的拓扑结构和学习算法进行了深入的研究，特别的还针对BP网引入RPROP(弹性BP)算法，对传统BP算法进行了改进；其次，本文论述了用Visual C++实现这几种网络的过程；文中还实现了BP网络的通用算法，建立了矿产资源综合评价及成矿预测BP模型，使用户可以自己创建、设计和管理成矿预测BP模型。 目前人工神经网络有三种常用网络模型：BP网、Hopfield网和Kohonen网。 BP网络(Back Propagation Networks，BP)是一种层状结构的前馈神经网络，具有非线性映射功能。由输入层、隐层和输出层组成，层内各神经元无连接，层间无反馈，信号沿同一方向从输入层经隐层传输至输出层。学习算法采用梯度搜索技术，使全局代价函数最小化，并据此调整连接权重，从而获取知识，并存储于层状网络的连接权中。它使网络具有知识结构严谨，推理机制高效和敏捷等特点。常用于进行函数逼近、模式分类、数据压缩等。 Hopfield网则是一种反馈网络，又称为自联想记忆网络，具有非线性动力学系统的许多优良特性。常见的结构为仅有一层神经元，各神经元间实现全反馈；其学习过程为设计一个网络，存储一组平衡点，使得当给网络一组初始值时，网络通过自行运行而最终收敛到所存储的某个平衡点上。此种网络主要用于联想记忆、模式分类、模式识别等。 Kohonen网是一种竞争式学习网络，它是模拟大脑神经系统的自组织特征吉林大学硕士学位论文瓢犷丙死可药而iion一瓦atur。Map)功能作为网络的学习方案，因此也称为SOM网络。其拓扑结构为:一个输入层，一个输出层，输出层节点以二维形式排成一个节点矩阵，输出节点之间可能实现局部连接，输入节点与输出层的所有节点通过权值实现全互联;它的基本思想是网络竞争层神经元竞争对输入模式的响应机会，最后仅有一个神经元成为胜利者，并对那些与获胜神经元有关的各连接权朝着更有利于它竞争的方向调整，这一获胜神经元表示对输入模式的分类。常用其进行无模式分类、聚类分析、优化设计等。 BP网是人工神经网络中应用最为广泛的网络模型，也是人们研究得最多的一种模型，在最基本的误差逆传播学习算法的基础上，人们又陆续提出了各种优化的算法，如基于全局速率调整的加入动量项、渐进自适应等方法和基于局部学习速率调整的符号变换等方法，以满足不同应用的需求。本文针对地质勘察资料数据的结构特点及其分类要求，对BP网进行了改进，使其更好的满足在地学中的应用。对BP网的改进主要有以下几点: 1.引入弹性BP算法对网络权值和闽值进行自适应修正，以克服传统算法中固有的学习收敛速度慢、容易陷入局部极小等问题。 2.对BP网来说，隐层节点数的选择对网络的性能影响很大，但如何选择目前并没有理论的指导，也没有好的解析式来表示。数目过多会使学习时间变长，数目过少会导致网络不强壮、识别能力差。为解决这一问题，我们在网络结构上也进行了适当调整，如各层神经元数目在学习过程中可随意设定，隐含层神经元的激活函数的类型也可自由选择，这样无论何种样本，通过多次调整隐层神经元数目和改变激活函数的类型，总能找到一个最优(稳定)的网络模型和最快的学习速度，提高了系统的通用性。 3.对输入样本进行归一化处理，使得过大或过小的样本输入值不至于令神经元过于饱和或截止，而恰好能落在神经元转移函数梯度最大的那些区域，保证学习能够收敛。 文中实现了BP网络的通用算法，建立了矿产资源综合评价及成矿预测BP模型，从而可以对成矿信息进一步进行智能化知识发现和信息挖掘，自动评估各地质变量对成矿的贡献，得到区域性的成矿规律和成矿模式，并圈定出成矿靶区。并提供友好的人机交互式界面，使用户可以自己创建、设更多还原

【Abstract】 This article was completed based on the accomplishment of the algorithm design sub-theme of the National Geologic Experiment Center’s 2002 project of the multiplicate geologic information processing technology based on GIS.It mainly investigated the computer programming realization of the mathematic model and computation method of the artificial Neural Network, and thus provided technical support to the nonlinear conformity management of the complex geological information.ANN( Artificial Neural Network) is a sort of large-scale self-adaptive nonlinear kinetic system, it can realize nonlinear mapping, mode identifying, function approximating, clustering analyzing, data compressing and design optimizing etc, and in the mean while, it has many favorable strong characters such as stability, astringency and robustness etc, so it has broad applications in data management area.There are lots of nonlinear problems in geological working. The integrated analyzing and sorting toward geological proved data, the accurate forecasting of the reserves and distribution of the mine resource etc all come down to the multiple data conformity disposal. Therefore, it is vitally important to develop a perfect nonlinear conformity disposal method. However, some favorable characters of ANN can content these geological working demands perfectly, so more and more people alter one after another their working focus to the nonlinear mathematic model based on the ANN in order to resolve the complex problems in geology much better.This kind of investigations have acquired some accomplishments at present, but much of these accomplishments were limited to resolve certain frondose application using certain fixed network model and were far from satisfying the multiple demands of geological proving. The purpose of this theme is to establish a multi-net-model and make it generally applicable during the arithmetic realization process to content all application demands.This article firstly made a in-depth investigation toward the network topo-structure and studying arithmetic of the three models of the ANN- BP Network, Hopfield Network and Kohonen Network, especially induct the RPROP(Flexible BP) arithmetic aiming at BP Network and made a improvement toward traditional BP arithmetic. Secondly, it discussed the process of using Visual C++ to realize these networks; Finally,a BP model for metallogenic prognosis can be constructed as the general algorithm of BP network was programmed. Users can create, design and manage BP models for metallogenic prognosis by interacting with a computer through a user-friendly interface.BP Network(Back Propagation Networks) is a kind of layer structured front-feed neural network with nonlinear mapping function. It is consisted of input layer, concealed layer and output layer, there is no connection between the nerve cells within the layers and no feedback between the layers, signals transfer along a samedirection from the input layer through the concealed layer to the output layer. Grades searching technology was adopted by the study arithmetic to minimize the public expense function and adjust the connection power by this, thus acquire knowledge and save into the connection power of the layer network. This endowed the network with the characters of preciseness knowledge structure and efficacious consequence mechanism. It was usually used to make function approaching, model sorting, data compressing etc.Hopfield Network is a kind of feedback network also called self associative memory which many favorable characters of nonlinear dynamics system. Its usual structure only has one layer of nerve cell, the feedback between each nerve cell is entire. Its study process is design a network, save a group of balance point so that when giving a group of initial values, the network finally converge to a certain saved balance point by self execute. This network is used mainly in associative memorizing, mode sorting and mode identifying etc.Kohonen Network is a kind of competitive study network, it simulate the Self-Or更多还原