【作者】 张绍辉；

【导师】 丁康；

【作者基本信息】 华南理工大学 ， 车辆工程， 2014， 博士

【摘要】 机械装备是一个复杂的非线性系统，由于振动信号包含着描述系统运行状态的丰富信息，振动信号分析是目前常用的机械系统故障诊断方法之一。一般通过在机械构件的关键部位测取振动信号，通过信号处理获得表征机械运行状态的特征指标，常用的信号处理方法有时域分析、频域分析及时频分析。但是实际测得的信号往往存在着非线性、非高斯分布的特点，加上故障形态的多变性，使得传统的信号处理方法在复杂系统故障诊断中存在着一定的局限性，有必要探索用于复杂非线性系统的故障诊断新方法，以防范于未然，减少损失。流形学习算法是近年来模式识别领域的研究热点之一，其本质在于通过一定的非线性映射将高维空间的数据结构在低维空间中进行表示，同时最大程度的保留高维空间数据的有用信息。因此，可以利用流形学习的优点，用以处理从时域、频域以及时频域中提取的多维信号特征或者处理由多个传感器获取的多源信号，实现机械运行状态的识别。然而，研究中发现存在以下问题：(1)噪声直接影响着流形学习算法的稳健性；(2)参数选择影响了算法的特征提取效果；(3)某些算法对高维数据结构信息保持不完整等。为此，本文从流形学习算法的基础理论出发，研究算法在机械系统状态识别、趋势分析中的噪声敏感性以及参数选择等问题，具体工作如下：(1)传统时域降噪方法需要消耗大量的计算时间及存储空间，不利于实现机械系统在线诊断。提出直接对特征样本空间进行降噪的方法，理论分析了进行特征空间降噪的可行性，试验结果表明所提方法可有效的降低计算时间，且明显提高机械运行状态识别及聚类精度，提高流形学习算法在机械故障诊断的适用性；(2)针对局部线性嵌入算法LLE(Local Linear Embedding)中近邻点数选择对降维效果影响非常敏感的问题，通过分析得出不同样本的最优近邻点数应该不相等的结论，进而提出了可变近邻的LLE算法，提高了算法的聚类效果。将LLE算法的泛化形式： NPE(Neighborhood Preserving Embedding)算法与自组织映射SOM(Self-Organizing Map)结合，实现轴承退化过程的状态识别；(3)针对局部保持投影LPP(Locality Preserving Projection)算法只考虑样本邻域信息而忽略距离较远样本信息的问题，提出同时考虑样本近邻信息及最远样本信息的保持投影算法：NFDPP(Nearest-Farthest Distantce Preserving Projection)，更好的保留数据结构的有效信息。发动机失火实验及轴承障实验结果表明，所提算法可有效提高机械运行状态的识别正确率；(4)针对谱回归算法SR(Spectral Regression)未综合考虑样本局部及全局信息的问题，提出了同时考虑局部结构和全局数据结构的谱回归分析算法(Local and GlobalSpectral Regression, LGSR)。发动机实验及变速器故障实验表明，改进的谱回归算法能够获得更高的识别精度和聚类效果；(5)针对多传感器测量系统，在前述研究基础上分别提出多维度的NFDPP(Multi-NFDPP)与多维度LGSR算法(Multi-LGSR)。将算法分别应用于多传感器监测系统的齿轮故障检测及轴承退化过程的在线监测，结果表明，这些方法能够有效的预测故障的发生并确定故障出现的部位。更多还原

【Abstract】 Mechanical equipment is a complicated nonlinear system. Vibration signal analysis isone of the common methods for fault diagnosis, as it contains rich information to describethe mechanical running conditions. The signals are measured from the key parts of amachine, and then features are extracted to represent the running status. The widely usedvibration signal processing methods can be divided into: time domain analysis, frequencydomain analysis, time-frequency domain analysis. Due to the system complexity, the signalis always nonlinear and non-gaussian, which makes the fault diagnosis more difficult. It isnecessary to find new fault diagnosis method for complex nonlinear system diagnosis,keeping the accidents from the beginning and reducing the loss.Manifold learning is one of the most popular focuses in pattern recognition, which isto represent the nonlinear data structure by a nonlinear map from the high-dimension spaceto a low-dimension space, and remaining the most useful information in the subspacesimultaneously. Therefore, we can adopt manifold learning to process various featuresextracted from time domain, frequency domain, time-frequency domain or to analysismulti-source sensor signals to recognize the machine running states. However, there arestill some problems when using manifold learning:(1) noise influences the robustness ofmanifold learning;(2) the parameter affects the mapping result;(3) some methods cannotpreserve the most useful information. So, based on the basics of manifold learning, weinvestigated different manifold learning algorithms in machine condition recognition andperformance assessment, with regard to the noise effect. The main research is as follows:(1)It is time-consuming and memory-consuming to de-noise time signal traditionally,which also leads to difficulties in real time diagnosis. Based on the feature analysis, thenoise contained in the vibration data is transferred to the features. Thus, we de-noise thesefeatures directly to enhance the computational efficiency and conserve the memoryrequirements, which is beneficial to the application of manifold learning in machine faultdiagnosis;(2)Since Locally Linear Embedding(LLE) is very sensitive to the numbers of nearestneighbor, which affects the dimension reduction. Based on the analysis of sample neighborselection, we propose a variable k-nearest neighbor locally linear embedding (VKLLE)algorithm to improve the classification and stability. NPE (Neighborhood PreservingEmbedding) is a linear approximation of the LLE, which is developed for out-of-sample problem. In this paper, NPE and SOM (Self-Organizing Map) are combined to assess thebearing degradation performance;(3)The LPP only focus on local neighborhood information, which neglects the otherfurther samples. A novel method named NFDPP (Nearest-Farthest Distantce PreservingProjection) is proposed to explore data structure by considering a sample’s nearestneighbors and farthest samples at the same time. The experiment results of thebearing-defect classification and engine-fault diagnosis validate that the proposed NFDPPapproach achieves the good performance;(4)Because the SR (Spectral Regression) does not take the global structure intoaccount, a novel feature extraction algorithm, called local and global spectralregression(LGSR), is presented for fault feature extraction. Gear and engine faultexperiments results demonstrate that the LGSR can extract identity information formachine defect classification;(5)Based on the research on NFDPP and LGSR, two multi-way data processingalgorithms, denoted as the Multi-NFDPP and Multi-LGSR, are presented for processingmulti-sensor signals. The Gearbox fault detection experiments and bearing degradationassessment indicate that the proposed algorithms can effectively predict the occurrence ofdefect and find the defect location.更多还原