【作者】 李敏；

【导师】 熊诗波；

【作者基本信息】 太原理工大学 ， 机械电子工程， 2010， 博士

【摘要】 机械设备健康状态监测的运行振动信号通常是故障诊断的重要数据来源,使用这些数据进行故障诊断可以通过建立基于信号的线性或非线性模型实现,也可以直接从振动信号中提取故障特征。使用何种诊断策略最为有效,则要根据数据性质的判断加以确定。如果数据来自于明显的线性系统,则采用基于线性模型的诊断方法是恰当的;若数据是非线性的,使用基于非线性模型的诊断方法可以取得好的效果;若数据显示机械设备进入混沌振动状态,则要提取混沌特征进行故障诊断。因此,本文研究了某些非线性数据特征的检验方法,例如使用双谱分析检验数据的非线性特征,并用于检验齿轮箱振动数据的特性;使用Lyapunov指数定量描述混沌程度,并以振动筛为例,研究筛帮不同部位的混沌性强弱。此外,通过相关分析研究各部分振动数据的相关性,以制定诊断策略。提出几何-物理空间概念,把大的系统的所有数据按物理空间划分成小区域的数据集,实现物理分区。在聚类分析的规则下,对时间上不断扩展的数据集进行基于距离的分类,实现数据分区。使用主元分析将高维数据空间降维成低维数据空间,在保持原有有用信息量几乎不变的情况下,去除冗余信息,仅使用较低的维数和较少的数据量来表示原有数据,并给出仿真算例。本文讨论了线性系统模型与可用于非线性系统的神经网络模型。从神经网络模型的拓扑结构优化、输出数据的个数、延迟步数、隐层神经元的个数和激活函数等方面优化神经网络的辨识能力,即提高辨识精度与加快辨识速度。从而提高了用于故障诊断的神经网络模型辨识的实时性。提出模型确定性的概念:将辨识系统的谱特征作为目标特征,对权值矩阵行数据做盒状图分析。权值离群值愈少,谱特征确定性愈好,也就是说模型是确定的,可以代表一定时间内的系统区域。本文提出了基于虚拟响应谱序列的诊断方法。在辨识出的精确模型基础上,使用并行仿真,获得系统对不同幅值的虚拟正弦或脉冲激励的响应,分析得到的正弦系列响应谱图和脉冲系列响应谱图,即可得到系统的动态特征,也可用于诊断线性或非线性系统的故障,从而提高了故障诊断的可信度。将这种方法应用于工程结构的诊断,取得较好的结果。更多还原

【Abstract】 Operation vibration signals of mechanical equipment health monitoring are usually the important source of data for fault diagnosis. In order to diagnose faults via these data, linear or nonlinear model based on signal can be used. It can be used to extract fault features directly from the vibration signals too. The effective diagnostic strategy is determined in accordance with results of inspection for nature of the data. If the data comes from the obvious linear system, the diagnosis methods based on linear model are appropriate. If the data are non-linear, the diagnosis method based on nonlinear model can achieve good results. If the data show that mechanical equipment is into the chaotic vibration state, the chaotic features are extracted for fault diagnosis. Therefore, this paper studies some test method of nonlinear data, such as bispectrum analysis that is used to test the characteristics of gear vibration data. Lyapunov index is used to describe chaos quantificationally.Propose the concept of geometry - physical space. All data of the large system are divided into smaller regional data sets, and get the physical partition.Under the rules in the cluster analysis, the data set that expands with time is classed under distance-based classification.The high dimensional data space is reduced into low-dimensional data space using PCA , the original information content is almost unchanged , redundant data is dislodged , the small amount of low dimension data are used to represent the original data, and simulation examples are given.Linear system models and neural network models that applie to nonlinear systems are discussed. The recognition capability of neural network is optimized by considering topological structure optimization, item count of output data, delay step, item count of hidden layer neuron, activation function, etc. It is that identification accuracy and speed are improved. Thus Real time property of identifying neural network models is improved.The concept of model determinacy is provided. Spectrum characteristics of the identified system that is target feature, boxplots of weight matrix row data are analyzed. Spectrum signature presents best determinacy, when there are less outliers of weight values.In this paper a Fault diagnosis method based on virtual response spectrum sequences is proposed. The parallel simulation is performed based on identified accurate models to obtain response sequences correspond to different amplitudes of virtual sine or impulse excition. The series of sine response spectrum and pulse response spectrum can be used to analyze dynamic characteristics of system. The series of response spectra can be used for fault diagnosis of linear or nonlinear systems too, thereby increasing the reliability of fault diagnosis. This method has been applied successfully to the diagnosis of engineering structures.更多还原