【作者】 刘小平；

【导师】 徐桂云；

【作者基本信息】 中国矿业大学 ， 机械设计及理论， 2013， 博士

【摘要】 机械设备在线监测是企业安全生产、产品质量保证的关键。一方面机械设备结构、运行状态复杂难以建立准确的数学模型，另一方面设备运行状态数据量大，非线性度高、噪声干扰强、不确定等特性使得故障诊断比较困难。本论文借鉴机器学习、故障诊断、人工智能等理论和应用成果，对复杂机械设备的智能故障检测、诊断进行了深入研究，主要内容有：(1)复杂非线性、动态信号处理以及故障统计量构造研究。该方法使用希尔伯特-黄变换（Hilbert-Huang Transform，HHT）振动信号分解到感兴趣的子频带；然后使用HHT把子频带信号分解为多个内蕴模式函数(Intrinsic Mode Function, IMF)，根据IMF系数的邻域相关性去噪，基于信号能量准则消除虚假IMF；提出基于数据依赖KICA（Data Dependent Kernel Component Analysisn, DDKICA）获取描述过程特征的内蕴信息，给出经验特征空间的DDKICA模型选择准则；最后根据抽取的时频域特征分布使用支持向量描述(Support Vector Data Description, SVDD)构造新的统计量、确定置信度进行故障监控。研究表明该方法能够及时发现异常情况。(2)基于多尺度理论的振动信号去噪和故障特征提取。分析了形态梯度小波的多尺度特性及其特点，使用形态梯度小波对振动信号进行多尺度分解，对各层的细节系数进行软阈值降噪处理，然后进行信号重构；对降噪后的信号采用S-变换进行多分辨率时频分析，从S变换谱图中提取故障特征。仿真和实例证明该方法能有效提取故障特征，适合在线监测和诊断。(3)先进机器学习理论在提升机故障监控研究和应用。针对具有冗余、异构（heterogenous）和多尺度特性的高维数据集，本文提出多核正交局部鉴别分析和全局保持（Multiple Kernel Orthogonal Locality Discriminative Analysis with GlobalityPreserving, MKOLDAGP）维数约简算法。该方法不仅保证了低维特征空间与原始数据空间具有相似的几何结构，具有更好的鉴别特性，而且使得数据局部聚类概率密度近似服从高斯分布。最后给出基于GMM的故障监测和故障统计量，较好地克服了现有因非线性、非高斯特性而导致高斯混合模型（Gaussian Mixture Model,GMM）的故障监测性能下降问题。仿真实验表明了本算法可以有效抽取数据特征，有较强的故障检测能力。(4)不平衡数据集的v-NSVDD多分类研究。分析了多类支持向量数据描述(support vector data description,SVDD)算法存在的问题，提出一种新的不平衡数据v-NSVDD多分类算法。该方法基于不同类别样本间隔最大原理，较好地克服噪声和在野点的影响，提高了分类模型的泛化性能；通过样本加权的方法解决了不平衡类别样本预测精度低的问题，并在理论上给出了根据类别样本数量设置样本加权系数的方法。为实现多分类器拒判，防止因每个分类器的核函数参数不同而影响判决结果的准确性和可靠性，本文给出基于相对距离和K-NN规则相结合的多分类方法。使用Benchmark数据集。进行仿真实验，结果表明本算法能够获得较低的分类误差，能够有效处理样本不平衡问题。更多还原

【Abstract】 Large scale machines are widely used in industry. The safe and reliable operationscan effectively improve the economic and social profits. As a result, the online perfor-mance monitoring of manufacturing process is essential for ensuring process safety andthe delivery of high quality of products. Although the fault detection and diagnosis tech-nologies for machines have been extensively researched in recent years, some openproblems have not yet resolved. The reasons are as following: on the one hand, the largemachines consist of many complex and different components, operate in complex andpoor working conditions and strong noising background. Thus, it is hard to model themin mathematics; On the other hand, a large amount of data collected from machines ischaracterized with complex nonlinearity, nonstationary, strong noises and uncertainties,resulting in more and more difficult to manage, for example, process and understand thesystem structure, identification of process states, feature extraction, fault monitoring sta-tistics construction, decision function development and etc. Inspired form the advancedtheory and technologies of machine learning, fault detection and diagnosis, artificial in-telligence and manifold learning theory et al., Taking account into the above problems,we take the large scale hoist as research subjects and study the theory and applications ofintelligent fault detection and diagnosis in dissertation. The main work includes the fol-lowing:(1) Signal processing method for dynamic and nonlinear signal and the constructionof fault are studied.(1) Hilbert-Huang transform-Data dependent kernel independentcomponent analysis (HHT-DDKICA) method is proposed to denoise and extract featuresfor dynamic and nonlinear vibration signals, i.e. a new denoising algorithm based oncorrelations among neighboring intrinsic mode functions (IMFs) coefficients obtained byHHT is proposed, and the true IMFs describing the original signals are selected to elimi-nate spurious IMFs according to signal energy criteria.(2) DDKICA is presented to de-termine intrinsic information source from IMFs, and a model selection criteria in the em-pirical feature space is also given.(3) Support vector data description (SVDD) is adoptedfor fault monitoring, and new statistics and confidence limits are established. Hoist ma-chinery application shows the efficiencies of the proposed method.(2)The applications of multiscale theory in signals denoise and fault feature genera-tion are researched. The morphological gradient wavelet (MGW) is performed to elimi-nate signal noises. The gear vibration signals are decomposed into multiple scales byMGW, the detailed coefficients in each scale are processed using soft threshold de-nosingand the true fault signals are reconstructed by the processed wavelet coefficients. Mul- ti-resolution S transform is employed to analyzed the reconstructed vibration signals, thegear fault features with good time and frequency resolution can be extracted from thespectral graphs. Simulations show that the proposed method can maintain multiscleproperty, and does not involve the problem of negative frequency and cross frequencydisturbances, resulting to extract the gear fault features effectively.(3)The applications of advanced machine learning algorithms in fault monitoringand disgnosis are researched. Inspired from manifold learning and multiple kernel learn-ing theory, a novel dimensional reduction algorithm called multiple kernel orthogonallocality discriminative analysis with globality preserving (MKOLDAGP) is developed todeal with redundant and heterogeneous features. The local and global geometric structureof the projected data in low dimensional space through MKOLDAGP is consistent withthat of original data set in high dimension space, the extracted feature also include locallydiscriminative information and the distributions of the local clusters approximate Gauss-ian distribution, overcoming the problem that nonlinearity and non-Gaussian of data re-sult in unsatisfying performance of Gaussian mixture model (GMM). A two-stage itera-tive optimization algorithm is proposed to adaptively adjust the weights and parametersof kernel functions, and then the fault statistics of GMM are developed to online condi-tion monitoring machine. Finally, a case study of hoist illustrates the efficiency of theproposed algorithm on the information extraction and fault detection(4)The imbalancedv NSVDD for multiclass classification is developed. To re-duce the influences of noise and outliers and correct the optimization problem in Ref[199],new unbalance datav NSVDD algorithm for multiclass classification is pro-posed. The samples are weighted to deal with the unbalanced data classification problem,and the method to determine the weight coefficients are deduced in theory. The proposedalgorithm can be extended to nonlinear cases by means of kernel trick. Most multiclassclassification methods are lack of rejection decision and the distances between the sam-ple and hypersphere centers in different feature spaces are not equal to their true distanc-es, which influence classification performance and reliability. Combining relative dis-tance with K-NN rule, a multiclass classification method is presented to deal with theabove problems. The results of benchmark testing show that the proposed method canprovide lower classification errors, overcoming the unbalanced data problem.更多还原