【作者】 杨昌昊；

【导师】 竺长安；

【作者基本信息】 中国科学技术大学 ， 精密机械与精密仪器， 2009， 博士

【摘要】 机械故障诊断是指应用测试分析手段和诊断理论,对机械设备运行中所出现的故障机理、原因、部位和程度进行识别和诊断,并根据诊断结论,进一步确定机械设备的维修方案或预防措施。机械故障诊断以状态识别为基础,从机械设备的异常状态出发,实现故障定位、定性及定因.但是在一般情况下,由于系统结构、运行过程以及外界环境的复杂性,机械设备的故障征兆与故障原因之间并不是简单的——映射关系,它们之间存在着复杂的一对多和(或)多对一关系,同时在诊断过程中还存在着许多随机的、模糊的、不确定的因素,从而导致在机械故障诊断过程中存在着大量的不确定性问题。传统的机械故障诊断方法如系统可靠性框图、故障树分析法等,难以解决机械故障诊断中的不确定性问题。而近年来快速发展的不确定性理论与方法,对于解决机械故障诊断中的不确定性问题具有很大的优势,并成为机械故障诊断领域中一个重要的研究方向。本文以国家自然科学基金“复杂工程系统故障预测与维护理论及关键技术研究”和国家863计划先进制造技术领域“十一五”重点项目“行业大型装备MRO支持系统”为背景,针对机械设备故障诊断过程中的不确定性问题,运用贝叶斯网络和证据理论,提出了一种新的故障诊断方案及相应的实现方法,并以转子系统为例进行了实验验证。本文的主要研究内容和创新性工作有:1)针对机械故障诊断中的不确定性问题,分析了故障诊断过程中不确定性的来源,提出了基于不确定性理论的会诊诊断解决方案.该方案利用专家的先验经验数据和(或)传感器数据建立相应的不确定性模型,并通过故障概率推理和多源故障诊断知识的集结实现快速、准确的机械故障诊断。2)为了解决机械故障诊断中的不确定性建模问题,提出了基于故障贝叶斯网络的不确定性建模方法。该方法通过建立故障树模型与故障贝叶斯网络模型之间的映射关系,实现了由故障树模型向故障贝叶斯网络模型的转化;针对故障树模型缺失的情况,提出了基于蚁群优化算法的结构学习方法,实现了由传感器数据构造故障贝叶斯网络模型。3)为了解决复杂机械系统故障推理中的NP难题问题,提出了基于联结树算法的故障概率推理方法。该方法包括贝叶斯网络结构转变、信念初始化、信念传递与吸收以及故障概率的计算等,其主要优点在于采用局部计算降低了故障概率推理的复杂度。4)提出了基于D-S证据理论的会诊诊断融合模型。该模型通过故障诊断实例与故障贝叶斯网络模型之间的贴合度分析,解决了多源故障诊断知识的融合悖论问题,实现了多个故障贝叶斯网络模型的诊断知识集结。5)以转子系统的故障诊断为例,分析了转子系统故障诊断过程中存在的不确定性问题,建立了转子系统的故障贝叶斯网络模型,并通过故障概率推理获得了可信的故障诊断结论。实验表明,上述方法切实可行。论文围绕机械故障诊断过程中的若干核心理论与技术问题进行研究,在智能故障诊断方法方面获得一些创新性的研究成果,为机械设备故障诊断提供了一种新的方法与工具。更多还原

【Abstract】 Mechanical fault diagnosis refers to the recognition and diagnosis of fault mechanism,fault causes,fault positions,and fault degree arising from the process operation by resorting to the means of test analysis and the theory of diagnosis.On these bases,maintenance plans and preventive measures could be further determined. Based on the abnormal state of mechanical equipment,mechanical fault diagnosis could realize fault location,make qualitative analysis,and fault reason analysis according to the state recognition.But in general,the mapping relationship between fault symptom and fault reason of the mechanical equipment is not just one to one,but many-to-one and（or） one-to-many because of the complexity of system structure, operation process,and running environment.Moreover,there exist multitudes of uncertainties in fault diagnosis process due to random,fuzzy,or uncertain factors.Traditional methods for making mechanical fault diagnosis such as System Reliability Analysis,Fault Tree Analysis could not solve uncertain problems existing in the mechanical fault diagnosis process.Uncertainty theories and methods developed rapidly in recent years,have apparent advantages in solving uncertainty problems,and become an important research field in mechanical fault diagnosis.The research in this dissertation is supported by the National Natural Science Foundation Project Fault Forecast and Maintenance Theories and Key Technologies of Complex Engineering Systems and the key project of the 11^th Five-year National Plan Industry Large-scale Equipment MRO Support System.By using Bayesian networks and D-S evidence theory,this dissertation put forward a new scheme and implementation method to solve the uncertainty problems in mechanical fault diagnosis,and made an experimental verification on the rotor system.The main research and innovation are as follows.1) On the analysis of the uncertainty source by considering its uncertainties in the process of mechanical fault diagnosis,this dissertation proposed a new consultation scheme based on uncertainty theories to solve its uncertainty problems. In the scheme,the uncertainty model was built by combining prior experience from experts with data from sensors,and the diagnosis results were obtained rapidly and accurately by probabilistic reasoning and knowledge integration.2) To solve modeling problems existing in the mechanical fault diagnosis,an uncertainty modeling method based on the fault Bayesian network was proposed.The transformation from fault tree model to fault Bayesian network model was realized by establishing the mapping relationship between them.Due to the lack of fault tree model under many circumstances,a structure learning method based on ant colony optimization algorithm was proposed.In this way,the fault Bayesian network model could be obtained from sensor da（?）a.3) A probabilistic reasoning method based on junction tree algorithm was proposed to solve NP hard problems existing in complex mechanical fault reasoning. The method consists of the transformation of Bayesian networks,the initialization of beliefs,the propagation of beliefs,and the calculation of fault probability,whose main advantage lies in the fact that the complexity of probabilistic reasoning is reduced significantly by using the local calculation.4) An information fusion model for consultation diagnosis was proposed based on D-S evidence theory.By analyzing the closeness degree between the fault case and the fault Bayesian networks,the model could be used to solve fusion paradox caused by different knowledge in the fault diagnosis.5) Taking rotor system as an example,Fault Bayesian Network model was built in this dissertation by analyzing the uncertainty problems existing in the fault diagnosis.A credible conclusion was drawn by resorting to the fault probabilistic reasoning.The numerical example indicates that the above-mentioned method is feasible.The dissertation conducted a study on a number of core theories and technologies in process of mechanical fault diagnosis and made some original research achievements in the intelligent fault diagnosis technology,which could provide a new method and tool for the mechanical fault diagnosis.更多还原