基于自适应PCA和时序逻辑的动态系统故障诊断研究

发布时间：2018-01-14 06:05

本文关键词：基于自适应PCA和时序逻辑的动态系统故障诊断研究　出处：《浙江大学》2017年博士论文　论文类型：学位论文

【摘要】：为了保证系统正常、安全地运行,故障诊断研究备受关注。通常,在复杂动态系统中,过程数学机理与知识难以预先获取,而过程数据很丰富且易于获取。因此,研究在纯数据模型框架下动态系统的故障诊断问题很有意义。针对动态系统故障诊断面临的模型自动更新和动态过程特征表达两大问题,本文分别从过程中全局和局部信息的自适应提取、模型的自适应稀疏化、时序逻辑的引入和扩展的角度,采用自适应主元分析(principal component analysis,PCA)和时序逻辑主要进行了以下几个方面的创新性研究:1、针对传统的PCA因模型全局、固定不能准确地反映动态时变和局部信息的缺陷,通过在每一种系统运行状况下对过程变量进行自适应分块,使得不同的运行状况有着各自的变量分块结果,并且充分利用块内局部信息和块间全局信息,根据每一个分块结果构建不同的全局和分块故障诊断模型,提出自适应分块PCA方法。最后探究该方法在TE过程中的应用,提高了故障检出率和故障分离的准确性。2、研究动态时变模型的自适应稀疏化问题。通过将不同的与各自系统运行状况相关的负荷向量稀疏惩罚项引入原始的PCA的优化命题的目标函数中,构建自适应稀疏PCA的优化命题,并且提出用于求解该优化命题的迭代内点算法,得到仅仅包含少量非零元素的稀疏负荷向量。然后,定义QT2和SPE统计量用于故障检测。如果发生了故障,则通过选取与故障相关程度大的变量并优先对它们进行故障重构的方式来实现故障分离。将该自适应稀疏方法应用在TE过程和波形系统的故障检测与分离中,验证了方法的有效性。3、探究纯数据驱动的复杂动态系统中过程时变特征的表达和学习问题,并且提出一种基于PCA和时序逻辑的多层故障检测与识别框架。利用时序逻辑以自然语言的形式表达系统中过程变量层、PCA所选取的主元层和监控统计量层的过程时变特征。提出一种新的分段线性拟合算法,用于统计学习时序逻辑公式。然后,在系统正常和故障运行状况下,分别创建正常和故障时序逻辑库。通过检查系统当前行为是否满足正常库中的时序逻辑公式来在线监测是否发生故障,如果发生故障,则基于故障库采用相似度测量来识别被检测故障的类型。最后将该多层故障检测与识别框架应用在TE过程中。4、针对现有的时序逻辑对动态时变特征表达力不足的问题,扩展一种新的时序逻辑,并将该扩展时序逻辑用于在线故障检测。首先,利用PCA选取其所提取的前几个主元作为过程特征变量。然后,根据这些特征变量,统计学习扩展的时序逻辑公式。最后,构建自动机形式的故障诊断模型,实现机器人手臂系统和波形系统的在线故障检测。
[Abstract]:In order to ensure the normal and safe operation of the system, the research of fault diagnosis has attracted much attention. Usually, in complex dynamic systems, the mathematical mechanism and knowledge of the process are difficult to obtain in advance. And the process data is rich and easy to get. It is significant to study the dynamic system fault diagnosis problem under the pure data model framework. There are two major problems in dynamic system fault diagnosis: automatic model updating and dynamic process feature representation. In this paper, the adaptive extraction of global and local information, the adaptive thinning of the model, and the introduction and expansion of temporal logic are discussed respectively. Adaptive principal component analysis (component) and temporal logic are mainly used in the following aspects of innovative research: 1. In view of the traditional PCA model because of the global, fixed can not accurately reflect the dynamic time-varying and local information defects, through each system in the operation of the adaptive process variables block. Make different running conditions have their own variable block results, and make full use of local information and global information between blocks, according to each block results to build a different global and block fault diagnosis model. An adaptive block PCA method is proposed. Finally, the application of this method in te process is explored, which improves the fault detection rate and the accuracy of fault separation. 2. In this paper, the adaptive sparse problem of dynamic time-varying models is studied. By introducing different load vector sparse penalty terms related to the performance of their respective systems into the objective function of the original PCA optimization proposition. The optimal proposition of adaptive sparse PCA is constructed, and an iterative interior point algorithm is proposed to solve the optimization proposition. The sparse load vector with only a small number of non-zero elements is obtained. Define QT2 and SPE statistics for fault detection. If a fault occurs. Then the fault separation is realized by selecting the variables which are highly correlated with the fault and giving priority to their fault reconfiguration. The adaptive sparse method is applied to the fault detection and separation of te process and waveform system. Verify the effectiveness of the method. 3, explore the expression and learning of process time-varying features in data-driven complex dynamic systems. A framework of multi-layer fault detection and identification based on PCA and temporal logic is proposed, in which the process variable layer in the system is expressed in the form of natural language using temporal logic. This paper presents a new piecewise linear fitting algorithm for statistical learning of temporal logic formulas. Then, under normal and fault conditions of the system, a new piecewise linear fitting algorithm is proposed for the process time-varying characteristics of the principal component layer and monitoring statistics layer selected by PCA. By checking whether the current behavior of the system is satisfied with the temporal logic formula in the normal library, we can on-line monitor whether or not there is a fault, if a fault occurs. Based on the fault database, similarity measurement is used to identify the type of fault being detected. Finally, the multi-layer fault detection and identification framework is applied in te process. 4. In order to solve the problem that the existing temporal logic is not sufficiently expressive to the dynamic time-varying feature, a new temporal logic is extended, and the extended temporal logic is applied to on-line fault detection. The first several principal components extracted by PCA are selected as process feature variables. Then, according to these feature variables, the extended temporal logic formulas are studied. Finally, a fault diagnosis model in the form of automata is constructed. The on-line fault detection of robot arm system and waveform system is realized.
【学位授予单位】：浙江大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：TP277

【相似文献】