基于Schur-ONPE的转子故障数据集降维方法研究

发布时间：2018-05-14 18:56

本文选题：故障诊断 + 数据降维　；参考：《兰州理工大学》2017年硕士论文

【摘要】：随着旋转机械故障诊断技术的不断研究与发展,它已经开始为人们所重视。而在工程实践中,随着设备的复杂度和信息量的增加,人们要获得的原始特征数据集变得越来越困难,呈现出信息量大、知识匮乏等问题。因此,如何从海量监测系统能采集的数据中去除不相关干扰与冗余信息十分重要,是当今故障诊断数据挖掘领域应该重点关注的问题。本项研究充分使用数据挖掘方法中的流形学习方法,开展对故障数据分类的降维研究。流形学习方法是一种能有效发现潜在于结构本质中信息的大数据驱动的方法。本研究的工作主要包含以下内容:1)将实验台上采集到的数据通过分析小波以及小波包的能量,筛选信号里存在的表现转子运行情况的信息,建立了相关数据集。此数据集能够排除一部分干扰信息,为后续工作的顺利展开奠定了基础。2)通过公式推导,比较分析主成分分析法(PCA)、核主成分分析法(KPCA)和邻域保持嵌入法(NPE)。采取实例验证的方式,对比得出邻域保持嵌入法在降维性能上的优越性。3)提出一种基于舒尔分解和正交邻域保持嵌入的降维算法,简称为Schur-ONPE算法。该算法运用舒尔分解替代了原本的正交邻域保持嵌入算法中的正交化迭代计算,削减了计算复杂度,有效提高了运算效率和准确程度。将Schur-ONPE算法的数据降维结果输入K近邻分类器之中进行的分类验证,发现得到的分类效果显著提高。再把不同转速下故障数据进行降维,也将结果输入到K近邻分类器之中,降维准确率也是稳定的,充分证明了该算法的有效性。4)将Schur-ONPE降维算法嵌入到LAB VIEW虚拟仪器技术和M ATLAB软件的混编程序中。结合了两种软件的优势,在原有的双跨转子实验软件平台上增加了经验模态分解模块和小波分析模块,进一步拓展了原有转子系统的软硬件功能,使该振动实验测试与反馈控制平台具有更好的人机交互性与信号处理能力。通过研究表明,利用数据挖掘算法,能有效挖掘出隐藏在海量检测数据背后的本质结构特征。进行创造突破性的研究,让更多的人认同并运用已经开展的研究,对该领域的智能化研究工作起导向性作用。
[Abstract]:With the continuous research and development of rotating machinery fault diagnosis technology, it has been paid more and more attention. In engineering practice, with the increase of equipment complexity and the amount of information, it becomes more and more difficult for people to obtain the original feature data set, which presents the problems of large amount of information and lack of knowledge. Therefore, it is very important to remove irrelevant interference and redundant information from the data collected by mass monitoring system, which should be paid more attention to in the field of fault diagnosis data mining. In this study, the manifold learning method of data mining is fully used to reduce the dimension of fault data classification. Manifold learning method is a big data driven method which can effectively discover the potential information in the nature of the structure. The work of this study mainly includes the following contents: 1) by analyzing the energy of wavelet and wavelet packet, selecting the information of rotor operation in the signal, the relevant data set is established. This data set can eliminate some interference information and lay a foundation for the smooth development of the subsequent work. Through formula derivation, the principal component analysis method (PCAA), the kernel principal component analysis (KPCA) and the neighborhood retention embedding method (NPE) are compared and analyzed. In this paper, the superiority of neighborhood preserving embedding method in dimensionality reduction is compared with that of example verification. (3) A dimensionality reduction algorithm based on Schuer decomposition and orthogonal neighborhood preserving embedding is proposed, which is called Schur-ONPE algorithm for short. The Schuer decomposition is used to replace the orthogonal neighborhood preserving embedding algorithm, which reduces the computational complexity and improves the efficiency and accuracy of the algorithm. The dimensionality reduction results of Schur-ONPE algorithm are input into the K-nearest neighbor classifier to verify the classification results, and the classification results are found to be significantly improved. Then reduce the dimension of the fault data at different rotational speeds, and input the results into the K-nearest neighbor classifier. The accuracy of dimension reduction is also stable. The validity of the algorithm. 4) embed the Schur-ONPE dimensionality reduction algorithm into the LAB VIEW virtual instrument technology and the mixed program of M ATLAB software. Combining the advantages of the two kinds of software, the empirical mode decomposition module and wavelet analysis module are added to the original two-span rotor experimental software platform, which further expands the hardware and software functions of the original rotor system. The vibration test and feedback control platform has better human-computer interaction and signal processing ability. The research shows that using the data mining algorithm, we can effectively mine the essential structural features hidden behind the massive detection data. Creative breakthrough research is carried out so that more people can identify with and use the existing research to play a leading role in intelligent research in this field.
【学位授予单位】：兰州理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP311.13;TH17

【参考文献】