多源信息融合中连续变量离散化及权重分配算法的研究

发布时间：2018-05-31 18:59

本文选题：多源信息融合 + 贝叶斯网络　；参考：《山东大学》2017年硕士论文

【摘要】：随着多传感器感知技术、智能计算技术以及无线通信技术的飞速发展,军事领域率先提出了"数据融合"这一概念,也就是将各式各样的传感器所采集到的信息加以"融合处理",进而能够获得比单个或单一种类的传感器更加行之有效的信息,因此对于多源信息融合技术的研究得到了社会的广泛关注。本论文主要进行了以下三个方面的研究:1.多源信息融合中连续变量离散化算法的研究本论文以基于概率论的贝叶斯网络数据融合算法为核心,针对现有算法对连续变量的离散化处理不够精确的问题,提出基于等期望值标准状态概率分配的连续变量离散化算法,更加科学合理地实现了对连续变量的无损处理,该算法解决了现有离散化方法的离散效果不够精确的问题。2.多源信息融合中多属性权重分配算法的研究多属性信息融合是多源信息融合的主要内容。现有的融合算法不可避免地遇到了各属性对于决策影响力即权重不同的问题,目前大多数权重分配算法或太过依赖专家知识或太过注重变量的统计特性,缺乏理论依据,只是定性地分析各连续属性的相对影响力的大小已不能满足现代数据融合定量分析的要求。本论文提出了一种基于信息增益的多属性权重分配算法,主要依据信息论,利用信息增益定量地计算权重,解决了现有算法定性分析所导致的权重计算不够精确的问题。3.加入权重的贝叶斯网络算法研究贝叶斯网络基于概率论,其参数学习方式是可监督式的。为了得到更加准确的推理结果,本论文采用了两种加入权重的方式:一是特征级的权重加入,即将权重加入到证据信息当中,进而得到权重化的证据信息后再进行网络推理;二是决策级的权重加入,由于贝叶斯网络支持不完备证据的特性,所以我们可以将各参数变量单独加入贝叶斯网络进行推理,最后将权重加入到它们的推理结果之中。仿真分析证明,加入权重能够有效地提升贝叶斯网络的推理准确率。通过对基于等期望值标准状态概率分配的连续变量离散化算法的研究,本论文实现了连续变量的无损处理,同时也将贝叶斯网络的应用环境拓展到了连续变量的推理融合之中。通过对基于信息增益的多属性权重分配算法的研究,本论文为权重分配带来了理论依据,能够实现定量地计算多连续属性的权重。通过对加入权重的贝叶斯网络算法的研究,本论文能够有效地提高贝叶斯网络推理的准确率,为贝叶斯网络的应用推广提供了坚实的基础。
[Abstract]:With the rapid development of multi-sensor sensing technology, intelligent computing technology and wireless communication technology, the concept of "data fusion" has been first put forward in the military field. That is, to "fuse" the information collected by a variety of sensors, and thus to obtain more effective information than a single or a single type of sensor. Therefore, the research of multi-source information fusion technology has been widely concerned by the society. This paper mainly carries on the following three aspects of research: 1. Research on discretization algorithm of continuous variables in Multi-source Information Fusion; this paper focuses on Bayesian network data fusion algorithm based on probability theory, aiming at the problem that the existing algorithms are not accurate in the discretization of continuous variables. A discretization algorithm for continuous variables based on standard state probability assignment of equal expectation value is proposed, which realizes the lossless treatment of continuous variables more scientifically and reasonably. The algorithm solves the problem that the discretization effect of existing discretization methods is not accurate enough. Research on Multi-attribute weight allocation algorithm in Multi-source Information Fusion; Multi-attribute Information Fusion is the main content of Multi-source Information Fusion. The existing fusion algorithms inevitably encounter the problem that the attributes have different influence on decision-making, that is, the weight is different. At present, most of the weight allocation algorithms either rely too much on the expert knowledge or pay too much attention to the statistical characteristics of variables, so they lack the theoretical basis. But qualitative analysis of the relative influence of each continuous attribute can not meet the requirements of modern data fusion and quantitative analysis. In this paper, a multi-attribute weight allocation algorithm based on information gain is proposed. Based on information theory, the weight is calculated quantitatively by using information gain, which solves the problem of inaccurate weight calculation caused by qualitative analysis of existing algorithms. Bayesian Network algorithm with weight; Bayesian Network is based on probability theory and its parameter learning method is supervised. In order to obtain more accurate reasoning results, this paper adopts two ways to add weight: first, the weight of feature level is added to the information of evidence, then the weight is added to the information of evidence, and then the weighted information of evidence is obtained and then the network reasoning is carried out. Secondly, the weight of decision level is added. Because Bayesian network supports the characteristic of incomplete evidence, we can add each parameter variable to Bayesian network separately and add weight to their reasoning result. Simulation results show that adding weights can effectively improve the reasoning accuracy of Bayesian networks. By studying the discretization algorithm of continuous variables based on the standard state probability assignment of equal expectation value, this paper realizes the lossless processing of continuous variables, and extends the application environment of Bayesian network to the inference fusion of continuous variables. Through the research of multi-attribute weight allocation algorithm based on information gain, this paper provides a theoretical basis for weight allocation, and can quantitatively calculate the weight of multiple continuous attributes. Through the research of Bayesian network algorithm with weight, this paper can effectively improve the accuracy of Bayesian network reasoning, and provide a solid foundation for the application and promotion of Bayesian network.
【学位授予单位】：山东大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP202

【参考文献】