【作者】 崔兆顺；

【导师】 汪志农；

【作者基本信息】 西北农林科技大学 ， 计算机应用技术， 2007， 硕士

【摘要】 目前,数据挖掘作为现代企业管理的重要辅助工具,被广泛应用于各个领域。在我国,数据挖掘技术取得了较大的成就,但还存在很多问题,尤其是在具体应用方面,面向实际应用的软件和可视化挖掘技术仍然十分匮乏,企业内各种信息往往不能有效的形成决策信息和知识,严重滞后了企业决策智能化的发展和应用。这也是造成我国企业尤其是西部中小企业的发展水平缓慢和管理模式落后的主要原因。所以急需一种能够解决现有问题,提高企业各种数据的处理能力,帮助企业进行辅助决策的应用系统。将数据挖掘技术应用于企业的辅助决策,发现数据中隐含的各种有用的知识,来提升企业管理和决策水平,是一种十分有效的手段。论文以西部地区中小型制造企业为研究对象,通过对日常时序数据进行模式分析,在国内专家知识和先进管理模式指导下,利用神经网络进行时序数据挖掘,用以达到销售量预测,为企业生产提供合理辅助决策,对企业的生产和发展具有一定的作用。论文系统研究了时序数据挖掘的相关理论和最新技术,总结了通用的开发方法。以此为基础,构建基于动态BP神经网络的时序数据预测平台,研究了其中所采用的关键技术并加以实现。论文主要研究成果有以下几个方面:(1)结合动态神经网络理论,提出了一种动态逐渐增元的数据预测框架,并以此建立了一个应用系统,在结构上,可以完成实际业务的数据挖掘相关工作,对时序数据进行有效的处理,发现数据中内在的规律和模式。(2)针对时序交易数据的特性,将预测转化为基于小样本空间的对应点误差最小预测,利用BP神经网络动态处理,以此来保证了数据挖掘后预测结果的精度保证;并在领域和专家知识的指导下,将其他经济指标参数作用于预测,提高预测的性能。(3)结合构建动态BP网络方法,进行动态可干预的数据预测,解决神经网络特征变量过多造成网络收敛速度降低的问题。利用可视化编程,结合灵敏度分析理论,提高了预测的精度和获得知识、规则的可解释和积累能力。论文的创新点在于利用动态神经网络和基于最小点误差预测应用于企业时序数据预测系统,提出了一种结合灵敏度分析和逐渐增元方法来提高预测结果解释能力的框架。利用将数据转化为小样本空间的方法,保障数据挖掘的精度。更多还原

【Abstract】 At present, data mining is the most important assistant tool of modern enterprises management,which be used to many fields. In our country,it has made great progress,but many problems are still in existence, especially in applicaton faced to specifical problems. Software of faced to application and data mining of visual technology are absent.The inner Information of enterprises can not come into being decision-making information and knowledge,which lags badlly to decision of intelligent, and which leads to low development and behindhand management pattern. So it is needed to provide an application system to solve the problem and improve capcibility for data processing, to help assistant decision for enterprises.The thesis aims to medium-minitype manufacture enterprise, processes the pattern analysis by time serials data.Under the knowledge of expert and advanced management, the neural network took the data mining’s tool to forecast sales volume,provider the reasonable decision-making for the production of the enterprise. Can solve the problems of data pre-processing and forcasting for medium-minitype enterprise, has certain effect to the production and development of the enterprise.The thesis systemiclly researched the theory and new technology bansed on time serials data mining, and summarized the general method. Based on it, the thesis built the platform of dynamic BP net, and reserached and realized the key technology.The followings are results of the research:(1) First, proposed the circulation dynamic neural network data forecast frame, and established a application system, in the structure, may complete the actual service for the KDD recurrent work, carries on effective processing to time series data, discovers the data intrinsic rule and the pattern.(2) In the view of the time series transaction data’s character, proposed predict transforms data based on the characteristic in smallsample space linear prediction, meanwhile had guaranteed after this the datamining predict the result in precision. (3)Through constructs the recurrent dynamic BP network method carries on forecast, which the dynamic recurrent may intervene, solved network characteristic variable excessively many caused the problem, which network convergence rate reduced. Then use visible programming, union Sensitivity analysis theory, increased the forecast precision and obtains the knowledge, the rule’s explanation and the accumulation ability.Finally has produced face terminal user’s concrete application system, enables to complete the data processing and the pattern discovered, and facilitates effectively carries on the data forecast. Confirmed in practical application has used the method’s efficiency and function. Because the forecast is the extremely complex system, for guarantee model accuracy and solution above question, therefore utilizes the mathematics theoretically analysis and the statistical axiom principle establishment correct data model in this article, and has carried on the confirmation through these principles and the theory to the system model.更多还原