【作者】 於东军；

【导师】 杨静宇；

【作者基本信息】 南京理工大学 ， 模式识别与智能系统， 2003， 博士

【摘要】 近年来，神经网络和模糊系统在理论与实践上取得了令人瞩目的进展。但是在神经网络和模糊系统的实际应用中，仍存在一些噬需解决的问题，本文对其中常见的几个问题进行了研究。 对于常用的三层结构的神经网络，隐节点数目的确定一直是个难题，至今无定论。本文研究了工程上常用的三层B样条神经网络，采用构造性的方法证明了B样条神经网络的全局逼近能力，并且给出了构造几乎最小隐节点的算法，从而在理论上为B样条神经网络的使用提供了依据。 针对呈现层次特征的应用领域，本文提出了相应的层次径向基神经网络(Hierarchical RBFN)，并且证明了HRBFN是一个全局逼近器。HRBFN更适合于具有层次结构的应用领域，并且HRBFN还能够部分消除使用RBFN会造成隐节点随着输入变量数目增加而急剧增多的问题。 模糊系统可以以两种方式应用于非线性系统辨识：串并联方式和并联方式。本文研究了串并联方式模糊系统的数字逼近特性，得出结论：当模糊规则数等于样本数时，已经可以实现精确插值，因此模糊规则条数不能超过样本数目，否则将冗余，并可能引起振荡，削弱模糊系统的泛化能力。此外还研究了系统逼近误差和初始状态误差对串并联模糊系统性能的影响，指出：只要模糊系统逼近实际系统足够好，即使两者之间存在初始状态误差，模糊系统仍能良好工作。 对于以并联方式进行非线性系统辨识时的模糊系统，本文研究了其预测收敛性和辨识收敛性。证明了只要并联模糊系统的参数满足一定的条件，可以保证并联模糊系统的预测收敛和辨识收敛，并给出了该条件。 应用模糊系统时，常常会遇到的一个问题就是系统中可能存在冗余的模糊子集和模糊规则。一方面增加了系统复杂性，浪费了计算能力，另一方面也给使用自然语言来描述系统造成了困难。本文针对TS模糊系统提出一种模糊子集和模糊规则的合并算法。使用该算法能有效地减少模糊子集和模糊规则的数目，进而减小了系统的复杂性，提高系统的可描述性。 最后，本文将神经网络和模糊系统用于人脸识别。本文先应用神经网络技术构建了一个完整的人脸识别系统：首先使用基于眼睛位置估计的方法从人脸图像中分割出对识别有意义的纯脸，然后使用自组织映射进行特征压缩，提取有效的鉴别特征，最后使用基于知识的模糊神经网络进行分类。 本文还提出一种人脸识别的模糊神经模型，该模型中的每条模糊规则用以估计模式在特征空间分布中的一个簇。通过自适应调节和模糊推理，对于每个给定摘要博士论文的模式，所提模型能够给出一个信度值表示该模式属于各个类别的程度。整个系统采用OCON的机构，该结构具有易收敛、识别快、适合分布式应用和增量式学习等特点。文中使用新颖的方法来确定模糊规则的条数和初始化模糊子集参数。实验结果证明该方法具有快的学习/识别速度、高精度和强鲁棒性的特点。更多还原

【Abstract】 In recent years, neural network and fuzzy system have acquired prominent successes both in theoretical and applied aspects. However, there still exists many problems need to be solved in real world application. Several common problems are researched in this paper.For the commonly used three-layered neural network, how to select the number of the hidden nodes is always a real problem. This paper researches the three-layered B-spline neural network, which is commonly used in engineering and presents a constructive algorithm for selecting the number of the hidden nodes. It is proved that the proposes algorithm can be used to build a B-spline neural network with minimum hidden nodes to approximate any continuous function defined on compact set to a prescribed accuracy.Considering some application fields may take on hierarchical characteristics, this paper proposes a corresponding hierarchical radial basis function network (HRBFN) and its universal approximation property is proved. Compared with the classical RJBFN, HRBFN is more suitable for application fields with hierarchical characteristics. In addition, HRBFN can partially solve the problem of the rapid increasing of the hidden nodes when the dimension of input increasing.Fuzzy system can be used in non-linear system identification in two modes: one is series-parallel mode and the other is parallel mode. This paper researches the numeric approximation characteristic of series-parallel fuzzy system and points out that the number of fuzzy rules should not exceed the number of the samples. In addition, the influence of approximation error and system initial error on the performance of the series-parallel fuzzy system is also investigated. As to the parallel fuzzy system, this paper proves that as long as the parameters of parallel fuzzy system meet some prerequisites, the parallel prediction procedure converges and the parallel identification algorithm locally converges.In applying fuzzy systems, a common problem is that there may exist redundant fuzzy subsets and rules, which can on the one hand increase the complexity of the fuzzy system and wasting the computational capacity, on the other hand make it difficult to describe the fuzzy system with natural language. In this paper, merging algorithms of fuzzy subsets and rules are proposed to deal with TS fuzzy systems.These two algorithms can effectively reduce the number of fuzzy subsets and rules, thus greatly decrease the complexity and enhance the descriptive characteristic of the fuzzy system.This paper proposes a fuzzy neural model for face recognition. The architecture of the whole system takes structure of one-class-in-one-network (OCON), which has many advantages such as easy convergence, suitable for distribution application, quick retrieving, and incremental training. Novel methods are used to determine the number of fuzzy rules and initialize fuzzy subsets. The proposed approach has characteristics of quick learning/recognition speed, high recognition accuracy and robustness. Experiments on ORL demonstrate the effectiveness of the proposed approach and an average error rate of 3.95% is obtained.更多还原