贝叶斯网在机械故障检测问题中的相关研究

发布时间：2018-03-24 19:38

本文选题：贝叶斯网络应用　切入点：机械系统检测　出处：《华中科技大学》2011年硕士论文

【摘要】：在现代机械系统的故障检测问题中，由于系统内部错综复杂的关系、信息测量手段的局限性、对系统知识的不甚了解等原因，会使得我们考察的问题本身具有较大的不确定性。贝叶斯网络作为基于概率论和图论的可视化网络模型，具有较强的自主学习能力和简洁直观的表达能力等诸多优越性，对于包含不确定性因素的复杂机械系统的相关问题研究具有很大的优势和广泛的应用前景。过贝叶斯网在具体应用中，也有很多问题需要考虑，比如样本过少，节点繁杂时，如何有效进行近似推理，贝叶斯网的节点赋值出现误差时我们怎么办，还有在应用贝叶斯网对机械进行故障检测时，需要安放许多传感器对系统进行信息读取，传感器过多可以更详尽地获取系统信息，不过过多的传感器会带有不少的“冗余信息”，并且会导致构建的网络底部节点相当多，加大网络学习成本。如何优化观察节点，提高推断效率是很有意义的工作。本文主要对贝叶斯网应用于机械故障检测中的上述关键问题进行一定研究和探讨，首先我们对实际应用中贝叶斯网的近似推理问题进行了研究，比对了两种随机模拟算法，，并指出利弊，以便在应用中更好的实施。其次，观察节点的测量误差在机械系统故障检测中较为常见，但是传统的方法象小波包去噪之类几乎全部是对于连续信息的去噪处理，本文引进了Gmbs抽样方法用于对于离散化后节点的信息去噪消除测量误差，进行了相关探讨，并期望在实际应用中有断推广。最后我们考虑在系统故障检测问题中构建的贝叶斯网络观察节点的简化问题，由于实际问题中经验信息的缺乏及对系统机理的不甚了解，使得我们安放的传感器接收了过多冗余的系统信息，从而导致观察节点过多，进而导致贝叶斯网络推断成本的加大，我们以汽轮机故障检测为实例探讨了贝叶斯网应用中观测节点的优化问题，结合常用统计手段主成份分析和因子分析对含有重叠信息的贝叶斯网的底部节点进行主要故障信息的提取，在呆留原有主要观察信息的基础上，简化贝叶斯网叶节点，构造新网络进行故障诊断，降低推断成本，提高推断效率。
[Abstract]:In the problem of fault detection in modern mechanical system, due to the intricate relations within the system, the limitation of information measurement means and the lack of understanding of the system knowledge, the problems we examine have greater uncertainty. As a visual network model based on probability theory and graph theory, Bayesian network has many advantages, such as strong autonomous learning ability and simple and intuitive expression ability. The research on the related problems of complex mechanical systems with uncertain factors has great advantages and wide application prospects. There are also many problems to be considered in the application of the Bayesian network. For example, when the samples are too small and the nodes are complicated, how to effectively carry out approximate reasoning, and what should we do when there are errors in the assignment of the nodes of the Bayesian networks? And when using Bayesian network to detect the fault of machinery, many sensors need to be put in to read the information of the system, and too many sensors can obtain the information of the system in more detail. However, too many sensors will have a lot of "redundant information", and will lead to a considerable number of nodes at the bottom of the network, which will increase the cost of network learning. How to optimize observation nodes and improve the efficiency of inference is a very meaningful work. In this paper, the key problems mentioned above in the application of Bayesian network in mechanical fault detection are studied and discussed. Firstly, the approximate reasoning problem of Bayesian network in practical application is studied, and two stochastic simulation algorithms are compared. And points out the advantages and disadvantages, in order to better implement in the application. Secondly, the measurement error of observation nodes is more common in mechanical system fault detection, but the traditional methods such as wavelet packet denoising are almost all for the continuous information denoising processing. In this paper, the Gmbs sampling method is introduced to eliminate the measurement error for the discrete node information denoising, and it is expected to be extended in practical application. Finally, we consider the simplification of Bayesian network observation nodes in the system fault detection problem, because of the lack of empirical information and the lack of understanding of the mechanism of the system. The sensor we put in receives too much redundant system information, which leads to too many observation nodes, which leads to the increase of the cost of Bayesian network inference. Taking turbine fault detection as an example, we discuss the optimization of observation nodes in Bayesian network application. Combined with principal component analysis and factor analysis, the main fault information of the bottom node of Bayesian network with overlapping information is extracted, and the leaf node of Bayesian network is simplified on the basis of retaining the original main observation information. A new network is constructed for fault diagnosis to reduce the cost of inference and improve the efficiency of inference.
【学位授予单位】：华中科技大学
【学位级别】：硕士
【学位授予年份】：2011
【分类号】：TH165.3;TP18

【参考文献】