【作者】 马君华；

【导师】 卢炎生；

【作者基本信息】 华中科技大学 ， 计算机软件与理论， 2010， 博士

【摘要】 数据挖掘(Data Mining,DM)是一种新兴的数据分析方法,它可以帮助人们充分应用数据中所蕴涵的信息,成为人工智能研究中非常活跃的领域。粗糙集是一种处理模糊和不确定知识的理论,聚类分析在没有先验知识时发现数据的规律,为人们提供了新的数据分类的方法。虽然在粗糙集和聚类分析方面有很多理论和方法产生,但由于数据对象的千变万化,需要我们不断对这些技术进行完善,以满足应用的需要。在研究粗糙集关联影响度的基本概念和性质的基础上,通过分析属性之间的相互影响,建立属性的关联影响度矩阵,以属性的关联影响度作为启发条件,有效地删除冗余属性,获得能反映出属性之间相互影响的约简集。实验表明,基于关联影响度的属性约简算法可以得到关联影响度较大的属性组成的约简集。这种概念拓宽了粗糙集的应用范围,为数据挖掘提供了新的方法。在粗糙集关联影响概念的基础上,对基于关联影响属性动态约简的概念和算法进行了研究,通过计算粗糙集中样本的激活状态ρ(U)和睡眠状态σ(U)对属性约简集的影响,在ρ(U)→σ(U)时,从约简集中将冗余的属性删除,在σ(U)→ρ(U)时,将必要的属性增加到约简集,这些算法和概念是有利于描述事物状态转换的方法。智能监控系统是工业自动化控制的核心,粗糙集理论为它提供了切实可行的实时决策规则,基于实时性的属性约简算法将弱实时性属性删除,保留强实时性约简集,以保证决策系统的实时性指标,该算法使粗糙集在实时决策系统的应用更加广泛。属性分类的约简算法能满足决策表中条件属性的分类要求,该算法按照分类函数对条件属性进行分类计算后,将次要的属性子集删除,求得属性分类约简集。实验表明,该算法能够在原决策能力不变的情况下,有效地删除部分属性,解决了属性分类的问题。将所研究的属性约简算法应用到配电网故障诊断和电网连锁故障诊断预警系统中。在分析和研究配电网故障诊断系统的属性选择和规则产生方法的基础上,将实时性属性约简、属性分类约简算法应用到配电网故障诊断系统中。通过电工理论计算电网连锁故障诊断预警系统的属性值,求出负荷转移情况下属性之间相互产生的关联影响以及线路的故障度。通过关联影响属性约简算法的应用,观察属性的变化,达到预测故障、及时排除故障的目的。调和聚类\分类算法,用于解决分类和聚类不一致的问题。它通过计算调和矩阵,计算聚类分类是否一致,通过对调和矩阵的不断修正,对聚类和分类的结果进行有效的协调,以达到最大程度上的一致。在电力负荷预测的应用中,该算法具有广泛的适用性,可以应用于其它分类和聚类不一致的场合。以上对数据挖掘进行的研究,经过模拟、试验和算例验证了算法的有效性,具有重要的理论意义和应用价值。更多还原

【Abstract】 Data mining is an innovated method of data analysis. It can help people maximize the useful information included in tremendous data, which has become active in artificial intelligence field. Rough set theory is a theory adopted to deal with rough and uncertain knowledge, which analyzes the clusters and finds the data principles when previous knowledge is not available, providing a new method for data classification. Although there are numerous methods of rough set and cluster analysis, as the data objects is changing continuously, we have to improve these relevant technologies over time, and propose creative theory in response, meeting the demands of application.This paper proposes the basic conceptions and attributes of relevant influences in rough set and study the interactions between different attributes, presenting a attributes reduction algorithm based on relevant influences. Through the matrix of relevant influences of attributes, making the relevant influences of attributes as inspiring prerequisite, we effectively delete redundant attributes to gain the reduced sets which reflects the interaction of different attributes. As proved by experiments, the algorithm could obtain the reduction sets composed of attributes with high relevant influences. This conception expands the application range of rough set, presenting a new method for data mining.Based on the conceptions of relevant influences of rough set, we study dynamic reduction conceptions and methods on the basis of relevant influenced attributes, and calculate the effects of activation stateρ(U) and dormancy stateσ(U) in rough set samples on the attributes reduction sets, whenρ(U)→σ(U), reduce the redundant attributes from reduction sets, whileσ(U)→ρ(U), add the indispensible attributes to the reduction sets, enabling the exchanges of event states be described more effectively by rough sets. This method compensates the deficiency of the previous methods that rough set could only describe static objects.As intelligence supervising system is the core of industrial automatic control, rough set theory has provided practicable real time decision principles, deducting the weak real time attributes and retaining strong real time ones, to ensure the real time principle of the decision system. The real time method of attributes reduction proposed by this paper has expanded the application of rough set in real time decision systems. As to the requirements of classification of conditional attributes in decision tables, this paper proposes a reduction algorithm on the basis of attributes classification, which first conducts classification calculation on conditional attributes according to classification functions, then deletes the less important subsets, concluding the classified reduction sets of attributes. As experiments have proved, maintaining the original decision ability constant, this algorithm could deduct parts of the attributes effectively and solves the attributes classification problems.We have applied our algorithm of attributes classification to the failure diagnosis of electric distribution network and early warning systems of electric interlock network. Studied the attributes choices and rules generation methods of failure diagnosis of electric distribution network, we employed real time attributes reduction and attributes classification reduction to the failure diagnosis systems of electric distribution network. Through calculation of the values of attributes of failure diagnosis and early warning systems of electric interlock network on the basis of electric theory, we have studied the relevant interactions between attributes under the condition of negative charges transforming, obtaining the extents of failures in the network. By the application of reduction algorithm of relevantly influenced attributes, we have observed the changes of attributes and achieved our goal of predicting failures and getting them fixed in time.Cluster/classification congruity algorithm, proposed in the situation of inconsistence between classification and cluster, is a method to achieve the results the accordance of cluster and classification respectively after calculates the congruity matrix, and coordinates the results effectively by modifying the matrix continuously, in order to achieve the maximum congruity. In the application of prediction of negative charges, this algorithm has a wide application, which could be used on some occasions when the results of cluster and classification are not consistent.These algorithm mainly conducts research on algorithms of attributes reduction and cluster analysis in rough sets, proposing several creative theories and methods, which also have been applied to electric automatic systems as experiments. As have been proved by experiments, these theories and methods are effective and practicable.更多还原