基于贝叶斯网络的气阀故障诊断研究

发布时间：2018-06-02 13:33

本文选题：气阀 + 故障诊断　；参考：《电子科技大学》2017年硕士论文

【摘要】：随着科学技术的不断发展,机械故障诊断技术越来越受到人们的重视。往复式压缩机作为典型的往复式机械,其内部结构复杂且激励源众多,传统的故障诊断技术已不能满足工程实际的需要。贝叶斯网络在处理不确定知识表达和推理方面具有独特的优势,已在语音识别、图像处理、金融分析等多个领域成功应用。因此,本文提出了基于贝叶斯网络的气阀故障诊断方法,该方法以气阀常见故障为对象,在研究了贝叶斯网络相关理论的基础上,从不同的角度建立贝叶斯网络结构学习算法,为其在故障诊断中的应用提供了有力的证据。文章在最后重点构建了两类贝叶斯分类模型,并将其成功应用于气阀故障诊断中。本文具体工作包含以下几个方面:1.阐述了贝叶斯网络的基本理论,并简单介绍了贝叶斯网络结构学习方法、参数学习方法和常见的四种贝叶斯分类器。2.针对气阀振动加速度信号,首先对原始信号进行小波阈值去噪,并通过小波包算法提取了各故障特征向量。将特征向量值与类变量值组成的样本进行离散化处理,将此作为贝叶斯分类器的输入。3.针对气阀常见故障,本文提出了一种BAN分类器算法。该算法首先利用遗传算法和K2算法构造属性节点之间的网络结构,然后加入这些节点的统一父节点(类节点)构造出分类模型,运用贝叶斯估计算法进行参数学习以获得各节点对应的条件概率表。根据测试样本集,以条件属性值作为证据,可求得测试样本的后验概率,最大后验概率所对应的类标签即作为该样本的分类结果。4.针对气阀常见故障,本文提出了一种GBN分类器算法。该算法首先利用CI测试去除与当前节点变量无关的变量,从而缩小了各节点的初始候选父节点集合的范围;通过贪心算法不断更新各节点的候选父节点,最终获得所求的分类模型。本文利用稀疏分数的方法进行故障特征选择,提取不同数量的特征集合,并利用GBN分类器进行分类预测。实验结果表明,通过该特征选择方法可以有效地提高气阀故障诊断正确率和减少计算的复杂度。5.总结了全文,并提出了下一步的研究方向。
[Abstract]:With the development of science and technology, people pay more and more attention to mechanical fault diagnosis technology. As a typical reciprocating machine, the reciprocating compressor has complex internal structure and numerous excitation sources. The traditional fault diagnosis technology can not meet the practical needs of engineering. Bayesian network has a unique advantage in dealing with uncertain knowledge representation and reasoning, and has been successfully applied in speech recognition, image processing, financial analysis and other fields. Therefore, this paper presents a method of valve fault diagnosis based on Bayesian network. This method takes common faults of air valve as an object, and establishes Bayesian network learning algorithm from different angles on the basis of studying relevant theory of Bayesian network. It provides strong evidence for its application in fault diagnosis. Finally, two kinds of Bayesian classification models are constructed and successfully applied to valve fault diagnosis. The specific work of this paper includes the following aspects: 1. The basic theory of Bayesian network is expounded, and the learning methods of Bayesian network structure, parameter learning and four kinds of Bayesian classifiers. For the vibration acceleration signal of the valve, the original signal is firstly de-noised by wavelet threshold, and each fault eigenvector is extracted by wavelet packet algorithm. The sample composed of eigenvector value and class variable value is discretized as the input of Bayesian classifier. In this paper, a BAN classifier algorithm is proposed for the common faults of the valve. Firstly, the network structure between attribute nodes is constructed by genetic algorithm and K2 algorithm, and then the classification model is constructed by adding the unified parent nodes (class nodes) of these nodes. Bayesian estimation algorithm is used for parameter learning to obtain conditional probability tables corresponding to each node. According to the test sample set, the posteriori probability of the test sample can be obtained by taking the conditional attribute value as the evidence. The class label corresponding to the maximum posteriori probability is regarded as the classification result of the sample. In this paper, a GBN classifier algorithm is proposed for the common faults of the valve. Firstly, the CI test is used to remove the variables independent of the current node variables, which reduces the range of the initial candidate parent node set of each node, and updates the candidate parent nodes of each node through greedy algorithm. Finally, the desired classification model is obtained. In this paper, the method of sparse fraction is used for fault feature selection, and different number of feature sets are extracted, and GBN classifier is used for classification and prediction. The experimental results show that this method can effectively improve the accuracy of gas valve fault diagnosis and reduce the computational complexity of .5. This paper summarizes the full text and puts forward the next research direction.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TH45;TP18

【参考文献】