基于模糊自动机的沉积环境判别方法研究

发布时间：2018-07-27 16:31

【摘要】：搬运介质、搬运方式、沉积环境和气候等因素控制着沉积物粒度参数的变化,因此,沉积物粒度分析对揭示气候变化和环境的演变具有重要意义。粒度参数中平均粒径、标准偏差、偏差和峰值是沉积物粒度的四个重要参数。不同粒度组分的形成与搬运均与沉积环境密切相关,处理、分析粒度数据有利于进一步确定沉积环境,这对于现代沉积学研究,乃至古代沉积物的沉积环境分析无疑都具有重要的理论和现实意义。模糊神经系统是这样一个模糊系统,它在处理数据样本过程中,使用某种学习算法去确定系统中的各项参数如模糊集和模糊规则,而该学习算法则由神经网络原理学习或激发出来的。利用模糊技术与神经网络的结合,我们可以充分发挥两者的优势,有效弥补各自的不足,这使得有关模糊神经系统的研究得到追捧,在包括自动控制、聚类分析和模式识别等领域中已得到成功应用,为人工智能的发展又增添了新的方向。本文是以诺姆·乔姆斯基提出的Chomsky体系文法为基础的,通过引入模糊自动机基本理论,简要概述了当前模糊自动机研究中有关神经网络技术的应用情况,这其中主要是以模糊有限状态自动机、模糊文法和神经网络这三者的关系为主线的,而通过对相关的神经网络进行训练,我们就可以抽取出所需的自动机,然后利用抽取的自动机进行文法的推导,以及对沉积环境进行判别分析。有关模糊神经系统的研究与应用已经具有一定的规模,但对于本论文来说,主要是研究基于神经网络的模糊正则文法推导算法以及模糊自动机的判别方法相关的问题。在模糊正则文法推导中,训练用于抽取模糊自动机的二阶反馈神经网络的算法主要有实时递归学习算法(RTRL)和实数编码基因遗传算法(RCGA),但是这两种算法存在通病：时间复杂度高,训练速度极慢。另外针对模糊有限态自动机(FSA)样本,两种算法都有不同种类的异常情况出现,也就是说泛化性很弱。最后,RTRL稳定性很弱；RCGA早熟现象经常出现。因而本文给出针对RCGA的改进算法：Levesbeg-Marguard Genetic Algorithm(LMGA),改进了训练速度以及解决了早熟的问题；针对RTRL算法,第七章给出了Levesbeg-Marguard Back Propagation(LMBP)算法,不仅加快了训练速度,增加了吞吐量,而且还能处理特殊情况下的字符串,比如超长串的情况。本文还给出了实验仿真与验证,并归纳了本文的结论。
[Abstract]:Factors such as transport medium, transport mode, sedimentary environment and climate control the change of sediment granularity parameters. Therefore, grain size analysis of sediment is of great significance in revealing climate change and environmental evolution. The average particle size, standard deviation, deviation and peak value of grain size are four important parameters of sediment granularity. The formation and transport of different granularity components are closely related to the sedimentary environment. Processing and analyzing granularity data are helpful to further determine the sedimentary environment, which is useful for modern sedimentology research. There is no doubt that the analysis of sedimentary environment of ancient sediments has important theoretical and practical significance. The fuzzy nervous system is such a fuzzy system, in the process of processing data samples, it uses some learning algorithm to determine the parameters of the system, such as fuzzy sets and fuzzy rules. The learning algorithm is studied or inspired by the neural network principle. With the combination of fuzzy technology and neural network, we can give full play to the advantages of both, effectively make up for their shortcomings, which makes the research on fuzzy nervous system become popular, including automatic control. Clustering analysis and pattern recognition have been successfully applied, which has added a new direction for the development of artificial intelligence. This paper is based on the Chomsky system grammar proposed by Norm Chomsky. By introducing the basic theory of fuzzy automata, the application of neural network technology in the research of fuzzy automata is briefly summarized. This is mainly based on the relationship among fuzzy finite state automata, fuzzy grammar and neural network. By training the related neural networks, we can extract the necessary automata. Then the extracted automata are used to derive the grammar and discriminate the sedimentary environment. The research and application of fuzzy neural system has a certain scale, but for this paper, it is mainly concerned with the derivation algorithm of fuzzy regular grammar based on neural network and the discriminant method of fuzzy automata. In the derivation of fuzzy regular grammar, the algorithms for training second-order feedback neural networks to extract fuzzy automata mainly include real-time recursive learning algorithm (RTRL) and real-coded gene genetic algorithm (RCGA),). The training speed is extremely slow. In addition, for the (FSA) samples of fuzzy finite state automata, the two algorithms have different kinds of abnormal cases, that is to say, the generalization is very weak. Finally, the stability of RTRL is very weak and the precocious phenomenon of RCGA often occurs. Therefore, this paper presents an improved algorithm for RCGA, called: Levesbeg-Marguard Genetic Algorithm (LMGA), which improves the training speed and solves the problem of precocity. For the RTRL algorithm, the seventh chapter gives the Levesbeg-Marguard Back Propagation (LMBP) algorithm, which not only speeds up the training speed, but also increases the throughput. It can also handle strings in special cases, such as long strings. The experimental simulation and verification are also given, and the conclusions of this paper are summarized.
【学位授予单位】：广东工业大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：P512.2;TP301.1

【参考文献】