【作者】 华斌；

【导师】 周建中；

【作者基本信息】 华中科技大学 ， 系统分析与集成， 2004， 博士

【摘要】 长期以来,我国水电机组的检修工作都是贯彻预防为主的方针,采取计划检修模式,定期进行检修。但由于水电机组各零部件之间使用寿命存在个体差异,同时其运行环境和外界影响也不相同,因此,按照某一固定的检修周期对水电机组进行计划检修,势必存在检修不足和检修过剩的现象。由此可见,我国水电机组传统的检修方式确实存在较多弊端。水电机组的状态检修是一种以水电机组工作运行状态为基础的预防维修方式,亦称预知或预测维修。它根据机组的状态监测和故障诊断所提供的信息,经过统计分析和数据处理,来判断机组的劣化程度,并在机组故障发生前有计划地进行适当的维修,能显著的提高水电机组运行的可靠性和降低水电机组检修费用。水力发电机组是一个复杂的非线性动力系统,其运行过程中故障的出现和发展包含大量的不确定性影响因素,而传统的故障诊断建模理论与方法难以对大量的不确定性因素进行精确的数学描述,导致实际的水电机组状态检修系统难以得出较为精确的诊断结论,极大的影响了水电机组状态检修在工程实际中的应用效果。为此,必须引入新的理论和方法,建立更加符合实际系统状态描述的有效模型,提高水电机组状态检修系统的实用性。贝叶斯网络是一种在复杂系统中建模、推理与机器学习的重要工具。它将数学中的概率理论与图论相结合,能很好的量化复杂系统中普遍存在的不确定性因素,其基于概率计算的推理方式能使得出的结论更加精确,同时基于贝叶斯网络的机器学习算法使建立的系统模型具有很好的智能性。近年来随着贝叶斯网络的理论与实现方法不断得到深入的研究,特别是基于贝叶斯网络的机器学习理论的完善,使得贝叶斯网络越来越多的被应用在各个领域,日益显示出其广阔的发展前途。应用贝叶斯网络的理论与方法,本论文系统、深入地研究了在水电机组状态检修工作中如何构建故障诊断专家系统与维修决策系统,提出了基于贝叶斯网络的维修决策与机组试验决策的一般决策策略,对贝叶斯网络的基本理论与方法进行了研究和探讨,并结合水电机组状态检修研究对贝叶斯网络的PPTC(Probability Propagation in Trees of Clusters)概率推理算法实现过程进行优化,最后,为解决贝叶斯网络在实际应用中存在的网络参数难以精确确定的难题,分别研究在完整数据集和不完整数据集下贝叶斯网络的参数学习问题,提出了在不完整数据集下进行参<WP=4>数学习的ME学习算法,使整个状态检修系统具有良好的自学习能力,为贝叶斯网络在水电机组状态检修系统中的深入应用进行了有益的探索。论文的主要研究内容包括:(1)研究水电机组状态检修实施的一般方法和步骤,提出了状态检修系统的一般体系结构,重点讨论了在状态检修系统中状态信号的采集、信号特征的提取以及机组状态的识别。(2)研究水电机组振动故障产生的机理以及具体的故障征兆,总结水电机组振动故障的传统诊断方法,根据振动故障的特点以及诊断的要求,提出了在水电机组状态检修系统中振动信号测点的布置方法。在贝叶斯网络PPTC推理算法的实现过程中,对其进行优化和改进以提高PPTC算法在反复推理时的执行效率。(3)首次将贝叶斯网络引入到水电机组的状态检修研究中,应用贝叶斯网络的理论与分析方法,以几种典型的水电机组振动故障作为研究对象,构造一个小型的水电机组故障诊断专家系统——SmartHydro。通过整个专家系统的建模、推理以及结论分析,对贝叶斯网络在实际中应用的理论与方法进行了系统完整的研究。(4)基于系统效用理论,提出了基于贝叶斯网络进行机组维修决策与试验决策的一般策略。作者通过对传统的贝叶斯网络图形结构进行扩展,在贝叶斯网络中引入新的节点类型,成功的在贝叶斯网络中实现了决策功能。通过具体决策任务的实现过程,清楚的表明借助于贝叶斯网络的理论与方法,能够真正做到故障诊断的结论为维修决策服务。(5)分别研究在完整数据集和不完整数据集下贝叶斯网络的参数学习问题,使水电机组的状态检修系统具有知识自动获取的能力。为了解决不完整数据集下参数学习的EM算法在某些情况下收敛速度较慢的问题,作者提出了一种基于最大信息熵的ME算法,并完整详细的介绍了整个算法的基本思想。最后通过实际应用中与EM算法的比较与对比,显示出ME算法在某些情况下所具有的优势。论文最后探讨并提出了今后水电机组状态检修研究工作的重点以及贝叶斯网络在水电机组状态检修中的研究方向。更多还原

【Abstract】 In the past, maintenance method of hydroelectric set often takes plan-maintenance mode, which is executed according to a fix period. But because there is individual difference in all kinds of parts in hydroelectric set, and environment and affect factor are different, the hydroelectric set maintenance method executed as a fixed period will bring insufficient maintenance and superfluous maintenance. So traditional maintenance mode have many shortcomings in fact. Condition-based maintenance is a kind of predictive maintenance mode, also be called Predictive-Maintenance. It can estimate the deterioration status of hydroelectric set by analyzing the data and the information provided by status monitor system and fault diagnosis system, and plan accurate maintenance schedule, it can promote reliability of hydroelectric set and reduce maintenance cost.Hydroelectric set is a complex nonlinear dynamical system, there are many uncertainty in appearance of fault, traditional modeling theory and method of fault diagnosis have difficulties in describing uncertainty accurately, which leads condition-based maintenance system of hydroelectric set have difficulties to get more precise diagnosis conclusion and to be applied into practice. So new theory and methods must be introduced to build more effective model and promote practicability of condition-based maintenance system of hydroelectric.Bayesian Networks is a kind of important modeling, inference and machine learning tool in complex system. It integrates probability theory and graph theory, it can perfectly quantizing uncertainty generally in complex system and can provide more precise result based on its probability inference, in the same time the system model based on Bayesian Networks has more intelligence. With more deeply research of Bayesian Networks of theory and implementation methods, specially machine learning methods of Bayesian Networks become more and more perfect, Bayesian Networks are applied into more and more fields and show its good future.The dissertation applies the theory and methods of Bayesian Networks into condition-based maintenance system of hydroelectric set, and study how to construct the fault diagnosis system and maintenance decision system, the general decision strategy of maintenance and test is put forward. In the same time as application research, the basic theory and methods of Bayesian Networks are introduced and discussed, and PPTC probability inference algorithm is improved. In the last, in order to solve the difficulties in <WP=6>really practice, ME learning algorithm under incomplete data set is brought forward, which leads the whole system has good self-learning ability and paves the road of applying Bayesian Networks into practice. The major content in this dissertation can be separated into some parts listed below:(1) Studying the general implementation methods and processes of hydroelectric set condition-based maintenance system. Putting forward the architecture of condition-based maintenance system. Discussing the signal collection, character extraction and status recognition in condition-based maintenance system.(2) Studying the vibration fault mechanism and the fault symptom in detail, summarizing traditional diagnosis methods of vibration fault, according to the requirement of fault diagnosis, presenting the laying methods of survey points in condition-based maintenance system.(3) Based on PPTC inference algorithm of Bayesian Networks, presenting a kind of optimized PPTC algorithm, and designing a kind of data structure to improve the performing efficiency of PPTC algorithm.(4) Applying the theory and methods of Bayesian Networks, with several typical vibration fault as study objects, Building a simple fault diagnosis expert system of hydroelectric set-SmartHydro. Through the modeling, inference and result analysis process of expert system, studying the application theory and method.(5) Based on theory of System Utility, presenting the general strategy of hydroelectric set maintenance and test decision. By extend更多还原