信号的稀疏表达在滚动轴承故障特征提取及智能诊断中的应用研究

发布时间：2018-01-19 07:33

本文关键词： 故障诊断滚动轴承智能诊断稀疏表达 l_1优化字典学习时频分布过完备结构稀疏流形学习图嵌入深度学习深度信念网络　出处：《中国科学技术大学》2017年博士论文　论文类型：学位论文

【摘要】：现代科学技术的迅猛发展让越来越多的先进技术融入于机械设备的故障诊断领域,并取得了令人惊叹的成绩,这使当今的故障诊断上升到了一个新的台面。稀疏表达技术作为新兴的信号处理方法,其自身优势十分适用于机械故障诊断。本文从稀疏表达理论的研究成果出发,探索了该方法对于机械故障信号的特征提取和故障诊断等方面的应用潜力。旋转机械在现代工业和生产中占有越来越大的比重。对旋转机械的运转异常做出及时预警不仅可以保证其运作的安全性,还可以带来明显的经济收益。滚动轴承作为实现其核心功能的关键部件在旋转机械中有广泛的应用,其运行状态的可靠性与否关系到整个机械系统的工作性能。因此,对于新的故障诊断方法的探索和开展,滚动轴承是一个很好的研究对象。基于上述考虑,本文从滚动轴承的故障诊断出发,基于稀疏表达理论提出了一系列的故障特征提取和故障诊断方法。主要内容如下:1.从历史背景、科技发展和实际工程案例等方面详细阐述了机械设备的状态监测和故障诊断这一选题的研究意义。以滚动轴承为研究对象,从其故障机理和信号特点出发,回顾并分析了现有的诊断研究方法,包括时域上的、频域上的和时频域上的诊断方法,并基于国内外的研究成果和研究现状讨论了各个方法的优缺点和当今故障诊断研究的不足方面。对智能诊断技术做了概述,并着重介绍了目前热门的神经网络和支持向量机技术。详细阐述了稀疏表达理论在故障诊断领域的三个主要方面的研究应用,即信号降噪、特征提取和故障分类,并在最后提出了本文研究的主要思路。2.对稀疏表达理论的基本概念做了详细介绍,从数学模型出发提出了稀疏系数求解和字典设计两个稀疏表达理论中的主要问题,并结合现有的研究状况对这两类问题分别详细阐述了几个常用方法且做了简单的比较分析。这些理论介绍为后续章节提出的诊断方法奠定了理论基础。3.针对传统小波变换的小波函数振动模式与机械故障信号的振动模式匹配不足这一缺陷,结合稀疏表达理论提出了一种可调小波振动模式的过完备小波变换。随后,以应用对比的方式分析了该变换相比传统时频变换用于特征提取的优势,并从中提取出可用于故障诊断的SWE特征。实验验证SWE特征相较以传统时频方法得到特征的分类性能更出色。4.利用结构稀疏表达理论建立起表征数据关系的结构字典并对各类信号进行有效表征,使得同类型的具有相似模式的滚动轴承振动信号具有统一的表达模式。系数求解过程中基于混合约束项进行计算优化,使得最终表达在组层面上和结构层面上均进行特征选择,准确把握信号的类别和结构模式。为了便于后续的分析诊断,从表达中进一步提出了低维故障特征SSW,并以实验证明该特征具有很强的噪音抑制能力和稳定性,可有效实现轴承的故障诊断。5.将数据的流形学习理论与稀疏表达理论相结合,针对滚动轴承故障诊断问题的解决提出了 ManiSC特征提取框架。该方法首先利用数据的先验知识建立起表征数据关系的图谱矩阵,再以流形学习方式找到数据的基矩阵,并将数据映射至稀疏域。实验证明了 ManiSC特征能以低维的方式有效表征出原始高维数据中的几何特性和内在数据结构,并比标准的稀疏表达和流形学习方法具有更出色的鲁棒性和特征辨识度。6.将稀疏表达与深度学习算法中的深度信念网络相结合(稀疏DBN),用其建立了分模块故障诊断网络,使一次诊断可以在评估出轴承故障位置的同时判断出故障严重程度。滚动轴承的寿命离散性特点会导致使用中的轴承存在先后失效状况,若不及时诊断则存在巨大的安全隐患。本文建立的分模块故障诊断网络能有效解决这个问题,且实验的对比分析说明利用稀疏DBN所搭建的诊断网络对于轴承的状态判断准确度极高,在工程上具有强大的应用潜力。
[Abstract]:The field of fault diagnosis of the development of modern science and technology makes more and more advanced technology in machinery and equipment, and achieved amazing results, the fault diagnosis of today's rise to a new table. The sparse expression as a signal processing method of emerging technology, its advantage is very suitable for mechanical fault diagnosis. This paper from the research results of the theory of sparse representation, explores the method for the potential applications of fault signal feature extraction and fault diagnosis of rotating machinery. Play more and more important role in modern industry and production. The abnormal operation of rotating machinery make a timely warning can not only ensure the safety of its operation, but also it can bring obvious economic benefits. The rolling bearing is a key component to realize the core function is widely used in rotating machinery, its running status can be Reliability relates to the performance of the entire mechanical system. Therefore, to explore new methods for fault diagnosis of rolling bearing and the development, is a good research object. Based on the above considerations, this paper from the rolling bearing fault diagnosis based on sparse representation theory, fault feature extraction and fault diagnosis method of a series of the main contents are as follows: 1.. Based on the historical background, science and technology development and practical engineering cases described in detail the condition monitoring and fault diagnosis of mechanical equipment on the significance of this research. Taking the rolling bearing as the research object, starting from the fault mechanism and signal characteristics, review and analysis of the diagnosis of existing methods. Including time domain, frequency domain of the diagnostic method in the frequency domain and time, and research results and research status at home and abroad are discussed based on the advantages and disadvantages of each method and the fault diagnosis. The lack of intelligent diagnosis technology are summarized, and emphatically introduces the current popular neural network and support vector machine technology. Elaborated the sparse representation theory in the field of fault diagnosis of the three main aspects of the research, namely signal de-noising, feature extraction and fault classification, and finally puts forward the main.2. thought this study on the sparse representation of the basic concepts of the theory in detail, from the mathematical model of the proposed sparse coefficients and two sparse dictionary design expression of major problems in the theory, combined with the existing research condition of the two kinds of problems are expounded, several commonly used methods and made simple comparative analysis. Laid the theoretical basis for the traditional vibration.3. wavelet transform and wavelet function vibration modes and fault signals of these theories introduced as diagnostic methods proposed in subsequent chapters Pattern matching overcomes the defect based on sparse theory proposed a wavelet adjustable vibration mode of overcomplete wavelet transform expression. Then, by way of contrast analysis of the transform compared to the traditional time-frequency transform for feature extraction, and extracted SWE can be used for fault diagnosis. The feature classification performance experiment verify the SWE characteristics and frequency characteristics of the traditional method are compared with the more excellent.4. structure using sparse representation theory to establish the relationship between the structure and characterization of data dictionary for effective representation of all kinds of signals, making the same type with a similar mode of rolling bearing vibration signal model with the unified expression. Based on the mixed constraint coefficient in the process of solving calculation optimization, making the final expression characteristics were carried out in the group level and structure level, accurately grasp the categories and structure of the model in order to facilitate the signal. Diagnostic analysis of follow-up, from the expression of the proposed low dimensional fault feature of SSW, and the experimental results show that the characteristics of strong noise suppression ability and stability, can effectively realize the bearing fault diagnosis.5. the data manifold learning theory and sparse representation theory are combined to solve the problem of fault diagnosis of rolling bearing is put forward ManiSC feature extraction framework. This method uses the data matrix of a priori knowledge of the established data representation of relations, and then to the manifold learning way to find the basis matrix data, and the data is mapped to the sparse domain. The experimental results show the characteristics of ManiSC with low dimensional effective characterization of geometric characteristics of the original high dimensional data and the internal data structure, and than the sparse expression and standard manifold learning method has better robustness and distinguishing features.6. learning algorithm and the depth of the sparse representation With the deep belief network (DBN, with its sparse) sub module fault diagnosis network is established, so that a diagnosis of bearing fault location and fault severity evaluation in life. The discrete characteristics of rolling bearings will lead to bearing used in the existing failure status after the first, if not timely diagnosis there is a huge security risk. This paper established the sub module fault diagnosis network can effectively solve this problem, and the comparative analysis of the diagnostic network using sparse DBN built for judging the state of bearing with high accuracy, has a strong application potential in engineering.

【学位授予单位】：中国科学技术大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：TH133.33

【相似文献】