【作者】 张德利；

【导师】 韩璞；

【作者基本信息】 华北电力大学（河北） ， 热能工程， 2008， 博士

【摘要】 在故障诊断领域,不确定性问题占多数,主要是由诊断对象的复杂性、测试手段的局限性、知识的不精确等原因导致的。特别是汽轮发电机组这样大型复杂的机电设备,其构件之间以及构件内部都存在很多错综复杂、关联耦合的相互关系,不确定因素和不确定信息充斥其间,其故障可能是多故障、关联故障等多种复杂形式。因此,解决不确定性问题是目前汽轮发电机组故障诊断中的首要问题。常用的解决不确定性问题的方法包括贝叶斯方法、粗糙集理论、证据理论等,经Agre G等多位专家的分析研究,发现基于贝叶斯理论的贝叶斯网络是目前解决不确定性问题的最有效的方法。本文研究了贝叶斯网络的最新发展,包括贝叶斯网络的表示、学习和推理。就贝叶斯网络推理进行了深入研究,提出了贝叶斯网络的简化推理算法。为克服简化推理算法占用内存的问题,提出了应用深度优先分支定界法以很小的时间代价换取较大内存空间,解决了简化算法的内存分配问题,具有很强的实用价值。以汽轮发电机组故障诊断中的不确定性问题为研究背景,系统地回顾了故障诊断的常用方法,阐述了汽轮发电机组常见的异常振动并对振动信号的频域特征进行了分析。提出了用于解决不确定性问题的故障诊断网络模型,并对模型的知识表达、建造方法进行了深入研究。提出了基于主成分分析方法与贝叶斯网络的汽轮发电机组故障诊断方法,采用主成分分析方法对易于提取的故障特征进行处理,从而获得初步的故障模式倾向,然后将获得的故障模式倾向作为贝叶斯网络模型的故障征兆节点,进一步进行诊断分析。该方法既避免了单独使用主成分分析法指标体系确定难的问题,又可以通过贝叶斯网络模型更好的把各种特征结合起来进行诊断分析,从而提高了诊断结果的可靠性。基于智能互补融合的思想将粗糙集理论与贝叶斯网络有机结合在一起,提出了一种汽轮发电机组故障诊断的新方法。利用粗糙集理论的知识约简技术对专家知识和故障特征进行压缩,去除冗余信息,获得最小诊断规则。同时利用贝叶斯网络来发现节点间的潜在关系,建立汽轮发电机组故障诊断的贝叶斯网络模型,在保证诊断结果准确率的基础上,缩减了冗余信息,从而提高了诊断效率。更多还原

【Abstract】 In the field of fault diagnosis, uncertainty problems are in the majority, mainly caused by the complexity of the target diagnosis, the limitations of the testing means, and imprecise knowledge. Especially for large and complex electrical equipment such as turbo-generator unit, there are a number of complex and coupling relationships among its components as well as themselves internal. Because of uncertain factors the faults may be multiple failures, associated faults or other complex forms. Therefore, how to resolve the uncertainty of turbo-generator fault diagnosis becomes the most important issue.Common methods of settling uncertainties problem include the Bayesian method, rough set theory, proof theory. Through Agre G and other expert research and analysis, Bayesian network based on Bayesian theory is the most effective method to solve the uncertainty problem.This paper studies the latest developments in Bayesian network express, learning and reasoning. The Bayesian network simplified inference algorithm is proposed after researching Bayesian network inference deeply. In order to overcome the memory occupied problem, a new optimizing algorithm based on depth-first branch and bound is proposed. It costs a little time for great memory space, solves the memory allocation problem. The method has a strong practical value.Take uncertainty problem of turbo-generator unit fault diagnosis as the research background, we systematically review the rotating machinery fault diagnosis methods and principles, expound the common turbo-generator abnormal vibration and analyze vibration signals in frequency domain characteristics. To solve the problem of uncertainty the fault diagnosis network model is proposed, and model of knowledge representation, method of construction is researched deeply.The turbo-generator fault diagnosis method based on the principal component analysis and Bayesian network is proposed. We using principal component analysis method to extract the easy handling fault symptoms and obtain the initial fault mode inclination, then the fault mode inclination will be fault nodes of Bayesian network, and further diagnostic analysis. The method avoids the principal component analysis indicator system identified difficult problem, and Bayesian network is better to integrate the various symptoms to diagnose, thereby enhancing the reliability of the diagnosis results. Based on complementary intelligent blending thought we combine rough set theory and Bayesian network organically, propose a new turbo-generator fault diagnosis method. Using knowledge reduced technology of rough set theory to compress expertise knowledge, remove redundant information, then obtain the minimum diagnosis rules. At the same time use of Bayesian network to discover the potential relationship among nodes, then build the turbo-generator fault diagnosis Bayesian network model. Further more it can raise diagnosis efficiency.更多还原