基于流形学习的滚动轴承早期故障识别方法研究

发布时间：2019-02-14 14:16

【摘要】：滚动轴承作为旋转机械装备中关键且易发生故障的零部件之一,其运行状态直接影响整个装备系统的性能,因此对滚动轴承进行状态识别与故障诊断研究具有重要意义。信号降噪和故障特征提取是状态识别与故障诊断中最核心的内容,由于滚动轴承运行环境复杂,干扰多、噪声大,且故障信号多为非平稳非线性信号,从而大大降低了传统降噪与故障特征提取方法的有效性。为此,本文以滚动轴承为研究对象,针对其强背景噪声下的早期故障振动信号特点,将流形方法与其它现代振动信号分析方法相结合,对滚动轴承故障信号的降噪和特征提取问题展开深入研究。论文主要内容如下： 1.论述了选题的背景意义及国内外滚动轴承故障诊断的发展进程；对当代滚动轴承故障诊断中的降噪和故障特征提取两个关键问题进行了详细综述；剖析了滚动轴承常见故障形式及其成因与后果；阐述了滚动轴承的振动机理及时域、频域、时频域振动分析理论的具体使用方法。 2.针对滚动轴承早期微弱故障特征易被噪声及干扰成分所淹没的问题,将KPCA流形、EMD和LTSA流形相结合,提出了改进EMD的早期微弱故障信号降噪方法。首先在EMD分解前先进行一次KPCA流形降噪,然后在EMD分解基础上将所有IMF系数利用LTSA流形算法提取其低维流形分量,并将其求和得到新信号,实现降噪。该方法不仅充分利用了EMD完全自适应性分析非平稳非线性信号的优势,还能有效克服噪声对EMD分解效果的影响,并很好地解决了常规EMD应用中忽略大部分分量造成故障信息丢失的问题。工程实际信号对比分析验证了该方法的有效性与优越性。 3.针对如何有效提取非平稳非线性故障信号敏感特征问题,提出了基于二维流形-Hilbert时频谱的滚动轴承时频故障特征提取方法。首先在Hilbert时频分析基础上,应用改进LPP算法的二维LPP流形算法提取信号流形成分,然后定义了奇异值熵定量衡量不同故障状态下流形特征的差异。该方法直接以二维信息为研究对象避免了一维流形算法需将二维信息转化为向量带来的信息损失,与一般PCA方法相比更能发现隐藏在高维数据流形结构中的局部数据特征。工程信号分析验证了该方法的有效性。流形奇异值熵与概率神经网络结合应用,进一步证实了该方法具有很高的可靠性。 4.针对工程实际操作与应用问题,开发了一套滚动轴承故障振动分析系统。将信号处理技术、故障诊断技术、数据库知识、虚拟仪器技术及人机交互技术相结合,采用递进式分模块进行研制,满足了工程现场方便快捷实用的需求。
[Abstract]:As one of the key and fault prone parts in rotating machinery equipment, the running state of rolling bearings directly affects the performance of the whole equipment system, so it is of great significance to study the status identification and fault diagnosis of rolling bearings. Signal de-noising and fault feature extraction are the core contents in state identification and fault diagnosis. Because of the complex running environment of rolling bearing, the disturbance is much, the noise is large, and the fault signal is mostly non-stationary and nonlinear signal. Thus, the effectiveness of the traditional methods of noise reduction and fault feature extraction is greatly reduced. Therefore, in this paper, the rolling bearing is taken as the research object. According to the characteristics of its early fault vibration signal under strong background noise, the manifold method is combined with other modern vibration signal analysis methods. The noise reduction and feature extraction of rolling bearing fault signals are studied in depth. The main contents of the thesis are as follows: 1. The background significance of the topic and the development process of rolling bearing fault diagnosis at home and abroad are discussed, and the two key problems of noise reduction and fault feature extraction in contemporary rolling bearing fault diagnosis are summarized in detail. This paper analyzes the common fault forms of rolling bearings and their causes and consequences, and expounds the vibration mechanism of rolling bearings and the concrete application methods of the theory of vibration analysis in time domain, frequency domain and time and frequency domain. 2. Aiming at the problem that the early weak fault characteristics of rolling bearings are easily submerged by noise and interference components, a new method for reducing noise of early weak fault signals based on improved EMD is proposed by combining KPCA manifold, EMD and LTSA manifold. First, KPCA manifold denoising is performed before EMD decomposition, then all IMF coefficients are extracted by LTSA manifold algorithm on the basis of EMD decomposition, and the new signal is obtained by the sum of IMF coefficients, which can be used to reduce noise. This method not only makes full use of the advantage of EMD complete adaptation to analyze nonstationary nonlinear signals, but also effectively overcomes the effect of noise on EMD decomposition. The problem of losing fault information caused by neglecting most components in conventional EMD applications is well solved. The effectiveness and superiority of the method are verified by comparing and analyzing the actual signals in engineering. 3. To solve the problem of how to effectively extract the sensitive features of non-stationary nonlinear fault signals, a novel time-frequency fault feature extraction method for rolling bearings based on two-dimensional manifold and Hilbert time-frequency spectrum is proposed. On the basis of Hilbert time-frequency analysis, the two-dimensional LPP manifold algorithm based on improved LPP algorithm is applied to extract the signal flow formation fraction, and then the singular value entropy is defined to quantitatively measure the difference of manifold characteristics in different fault states. This method takes two-dimensional information as the research object directly and avoids the information loss caused by the one-dimensional manifold algorithm which needs to transform the two-dimensional information into vectors. Compared with the general PCA method, it can find the local data features hidden in the high-dimensional data stream shape structure. The engineering signal analysis verifies the effectiveness of the method. The application of Manifold singular value entropy and probabilistic neural network further proves that this method is highly reliable. 4. A rolling bearing fault vibration analysis system is developed for practical operation and application. The signal processing technology, fault diagnosis technology, database knowledge, virtual instrument technology and human-computer interaction technology are combined to develop the progressive sub-module.
【学位授予单位】：大连理工大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TH133.33;TH165.3

【参考文献】