【作者】 张晓丹；

【导师】 赵海；

【作者基本信息】 东北大学 ， 计算机应用技术， 2005， 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。特别是汽车发动机这种大型复杂的机电设备，其构件之间及构件内部都存在很多错综复杂、关联耦合的相互关系，不确定性因素和不确定性信息充斥其间，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车发动机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、模糊集理论、证据理论等，经Agre G等多位专家的分析研究，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。 本文以发动机故障诊断中不确定问题为研究背景，以Audi A6 1.8L型发动机为应用研究对象，提出了用于解决不确定性问题的故障诊断结构模型和功能模型，并对模型的知识表达、建造方法进行了深入研究，提出了贝叶斯网络融合算法。为了获得更高准确率的故障诊断结果，针对故障诊断中传感器传输数据的时序性特点，采用在线学习的思想，对贝叶斯网络结构进行改进。以MLE方法进行参数学习。并通过图搜索算法，推理得出产生故障的节点，获得高效、准确的诊断结果。 在应用中，将故障诊断融合模型及融合算法应用在汽车发动机故障诊断实际过程中。通过多个实验表明，在发动机运行工况下，故障诊断融合模型的诊断准确率优于传统的专家系统方法和学习前的贝叶斯网络方法，有效地解决了故障诊断中存在的不确定性问题，提高了诊断的准确率。从而验证了故障诊断融合模型及算法的有效性和应用价值。更多还原

【Abstract】 Uncertain problems are key problems in fault diagnosis fields, which are resulted by many important reasons, including complex diagnosis objects, limit test means, and the inexact diagnosis knowledge, etc. Motor is large mechaatronics equipment, which has complex, correlative relation exiting in its units, and the units are full of uncertain factors and information. So the fault of motor is possible multi-kinds, correlative. In the all, the most important problem of solving motor fault diagnosis is to solve the uncertain problem.There are many common methods for solving uncertain problems, such as Bayesian method, fuzzy theory, DS theory, certain theory etc., And there are many experts researching on these methods, and has obtained that, Bayesian network based on Bayesian theory is the best method for solving uncertain problems now.In this paper, for solving uncertain problems, the fault diagnosis structure model and function fusion model was proposed, the models were based on the information fusion technology and the research of uncertain problems in fault diagnosis process, and the diagnosis object was Audi A6 1.8L motor. Bayesian network fusion algorithm was proposed for the knowledge expression and constitution of the models, which were researched deeply.Construct of Bayesian network included structure learning and parameter learning. There are many methods for structure learning. According to the information obtained by sensors timely, an improved structure learning method of Bayesian network was proposed in order to get more exact diagnosis result. The on-line learning idea was adopted for the construct of structure, and the learned structure should be close to the former structure, the MDL of learned structure should be less than the former structure. Then the data from sensors inputting into the system continuously, the new structure was obtained by learning algorithm continuously. For getting parameters of Bayesian network, MLE method was adopted, which is adapt to solving uncertain problems.更多还原