【作者】 张恒；

【导师】 李凌均；

【作者基本信息】 郑州大学 ， 机械电子工程， 2011， 硕士

【摘要】 在设备故障诊断中,传统的数据处理方法只能针对单通道数据进行分析,而单通道数据往往不能将设备的空间运动信息完整的表征出来。而作为全信息分析方法的一种,全矢谱分析技术在处理同源多通道故障信号的同时能够体现更全面准确的转子运动空间特征信息。在此基础上本文将全矢谱技术与支持向量数据描述相结合,提出了全矢谱支持向量数据描述(Vector Spectrum Support Vector Data Description, VSSVDD)故障诊断方法。针对支持向量数据描述(Support Vector Data Description, SVDD)分类方法中训练样本数目受限的问题,本文对SVDD分类器作二次改进,引入动态支持向量数据描述(DSVDD)分类方法。该方法在训练样本中不断注入新的样本进而不断更新分类边界,从而更准确的表征了目标样本的区域边界。本文主要研究和解决问题如下：第一,支持向量数据描述方法是建立在统计学习理论之上的,核函数的引入可以把低维空间的非线性问题转化为高维空间的线性问题。选择不同的核函数对SVDD的分类效果不同。第二,运用全矢谱分析方法对采样数据进行分析处理,并且提取典型倍频上的幅值作为SVDD分类器的特征向量。实验表明经过全矢谱特征提取后的SVDD的分类效果较未经特征提取SVDD的分类效果更为明显。通过实验研究验证了全矢谱支持向量数据描述故障诊断方法对测试样本进行分类的可行性与有效性。第三,运用全矢谱支持向量数据描述方法对设备性能退化评估引入隶属度和相对距离的概念避免了超球体边界误差带来的影响,可以将测试样本的状态更加精确的表述出来；同时又体现了状态变化的过程。第四,提出动态支持向量数据描述分类方法的改进型。该方法的提出改变了原来SVDD分类方法中,分类器经过一次训之后分类边界永不改变的现状。它将测试得到的目标样本与本次测试以前的支持向量集一起形成新的训练样本,然后对SVDD重新训练。这样分类边界将更能体现设备的正常样本特征。更多还原

【Abstract】 In the equipment fault diagnosis, the traditional methods of data processing can only analyze the data for the single channel, while the single-channel data often can not completely characterize out of the information on the equipment space motion. As one of the full-information analysis methods, the full vector spectrum analysis technology while dealing with homologous multi-channel fault signal, it can reflect a more comprehensive and accurate information about spatial characteristics of the rotor movement. Based on that, this paper will combine the full vector spectrum technology with support vector data description, and then propose the vector spectrum support vector data description (VSSVDD) fault diagnosis method. For the problem which support vector data description (SVDD) classification method is limited in the number of training samples, this paper presents the second modification of the SVDD classifier, Dynamic support vector data description (DSVDD) method. This method continue to update the boundary of classification by injecting new samples for the training samples and thus characterize out of the boundary regions of target samples. The problems which this paper studied and solved are as follows:First, the support vector data description method is based on statistical learning theory. The introduction of the kernel function can make the nonlinear problems in low-dimensional space transformed into linear problems in high dimensional space. Selecting different kernel functions makes the effect of SVDD classification different.Second, used full vector spectrum analysis method to analyze and process sampled data and extracted the amplitude on typical frequency multiplication as a feature vector of SVDD classifier. Experiments showed that the effect of SVDD classifier extracted by vector spectrum feature was more obvious than that without the extraction. Verified by an experiment, the vector spectrum Support Vector Data Description fault diagnosis method is feasibility and validity as to classifying the test samples.Third, used vector spectrum support vector data description method to introduce the concept.of the membership grade and the relative distance for degradation assessment of equipment performance to avoid the impact brought by the different boundaries of super-sphere of different, so that the state the test samples can be more represented more accurately; while reflects the process of the state changing.Fourth, proposed an improvement of SVDD, that is dynamic support vector data description. The raise of this method changed the current situation where after the classifier experiencing a training, classification boundaries never be changed in the original classification method of SVDD, It putted the goal test sample together with the support vector prior to this test to form a new set of training samples, and then re-trained SVDD. Thus training samples continuously updated classification boundary which also better reflect the characteristics of normal Status of the device.更多还原