基于主元分析的故障检测与诊断研究

发布时间：2018-06-19 11:28

本文选题：主元分析(PCA) + 小波变换　；参考：《南京师范大学》2011年硕士论文

【摘要】：由于现代化大生产的迅速发展以及科学技术的迅速进步,现代工业系统的结构越来越复杂,投资越来越大,自动化水平越来越高,因此系统过程的安全性和可靠性就显得特别重要。基于主元分析的多元统计故障检测与诊断技术是目前过程自动化和控制领域的研究热点问题之一本文首先介绍了故障诊断方法、多元统计方法和TE过程模型,研究了基于主元分析方法的故障检测技术以及利用贡献图进行故障诊断的基本方法。针对主元分析方法在进行故障检测与诊断时存在的不足,将主元分析与小波变换、RBF神经网络等方法相结合,分别提出了基于小波去噪主元分析的故障检测与诊断方法,基于鲁棒主元分析的故障检测与诊断方法以及基于RBF神经网络的非线性主元分析的故障检测方法,将动态主元分析用于故障检测与诊断,提出基于动态主元分析的故障检测与诊断方法,以TE模型为对象进行了仿真研究和仿真结果分析。主要研究工作为： 1、研究了主元分析的基本原理,简单介绍了田纳西-伊斯曼(TE)模型,给出了基于主元分析的故障检测与诊断算法流程,通过TE模型进行仿真研究,根据SPE和T2统计量的变化来判断是否发生故障,根据变量对统计量的贡献来判断故障变量、识别故障源,实现故障的检测与诊断。 2、针对传统主元分析在处理含噪数据时的不足,研究了小波变换的基本原理,结合基于主元分析方法的故障检测与诊断方法,提出基于小波去噪主元分析方法的故障检测与诊断方法,TE模型的仿真研究表明该方法能有效地减少主元个数,降低误报率,提高了故障检测与诊断的效果。 3、针对传统主元分析方法要求建模数据的噪声服从正态分布,提出了一种基于鲁棒主元分析的故障检测与诊断方法。该方法使用简单的加权方差-协方差的估计值代替传统的协方差,在此基础上建立主元模型构造SPE和T2统计量来检测过程故障并根据变量对统计量的贡献来判断故障变量、识别故障源,TE模型的仿真研究表明了该方法优于传统的主元分析方法。 4、针对传统主元分析方法不能有效地监视动态多元过程,提出了一种基于动态主元分析(DPCA)的故障检测方法,根据测量数据建立动态主元模型,在该模型基础上利用SPE和T2统计量进行故障检测,以TE模型为对象进行了仿真研究,证实了基于动态主元分析进行故障检测时考虑时序相关性是由于传统主元分析。 5、针对传统主元分析方法的非线性局限性,将RBF神经网络与非线性主元建模相结合,提出基于RBF神经网络的非线性主元分析故障检测方法,利用RBF神经网络训练学习得到非线性主元的负载矩阵从而建立主元模型,仿真研究表明该方法在对非线性系统进行故障检测时优于传统的主元分析方法。
[Abstract]:As a result of the rapid development of modern mass production and the rapid progress of science and technology, the structure of modern industrial systems is becoming more and more complex, the investment is increasing, and the level of automation is becoming higher and higher. Therefore, the safety and reliability of the system process is particularly important. Multivariate statistical fault detection and diagnosis based on principal component analysis (PCA) is one of the hot issues in the field of process automation and control. This paper first introduces the fault diagnosis method, multivariate statistical method and te process model. The fault detection technology based on principal component analysis (PCA) and the basic method of fault diagnosis based on contribution diagram are studied. Aiming at the shortcomings of principal component analysis in fault detection and diagnosis, combining principal component analysis with wavelet transform RBF neural network, a fault detection and diagnosis method based on wavelet denoising principal component analysis is proposed. The method of fault detection and diagnosis based on robust principal component analysis and nonlinear principal component analysis based on RBF neural network is presented. Dynamic principal component analysis is applied to fault detection and diagnosis. A fault detection and diagnosis method based on dynamic principal component analysis (DPCA) is proposed. The te model is taken as the object of simulation and the simulation results are analyzed. The main research work is as follows: 1. The basic principle of principal component analysis (PCA) is studied, and the Tennessee Eastman (TET) model is briefly introduced. The flow of fault detection and diagnosis algorithm based on PCA is presented, and the simulation is carried out through te model. According to the changes of SPE and T2 statistics to determine whether the fault occurs, according to the contribution of variables to the statistics to judge the fault variables, identify the fault source, 2. Aiming at the deficiency of traditional principal component analysis in dealing with noisy data, the basic principle of wavelet transform is studied, and the fault detection and diagnosis method based on principal component analysis is combined. A fault detection and diagnosis method based on wavelet denoising principal component analysis (PCA) is proposed. The simulation results show that the method can effectively reduce the number of principal components and the false alarm rate. The effect of fault detection and diagnosis is improved. 3. In view of the noise distribution of modeling data required by traditional principal component analysis method, a fault detection and diagnosis method based on robust principal component analysis is proposed. In this method, a simple weighted variance-covariance estimate is used to replace the traditional covariance. On this basis, the principal component model is constructed to construct SPE and T2 statistics to detect process faults and to judge the fault variables according to the contribution of variables to the statistics. The simulation study of identifying fault source te model shows that this method is superior to the traditional principal component analysis method. 4. The traditional principal component analysis method can not effectively monitor the dynamic multivariate process. A fault detection method based on dynamic principal component analysis (DPCA) is proposed. Based on the measured data, a dynamic principal component model is established. On the basis of this model, SPE and T2 statistics are used for fault detection. The te model is used as an object of simulation. It is proved that the consideration of temporal correlation in fault detection based on dynamic principal component analysis is due to traditional principal component analysis. 5. Aiming at the nonlinear limitation of traditional principal component analysis method, RBF neural network is combined with nonlinear principal component modeling. A nonlinear principal component analysis (NPCA) fault detection method based on RBF neural network is proposed. The load matrix of nonlinear principal component is obtained by training RBF neural network and the principal component model is established. Simulation results show that this method is superior to the traditional principal component analysis method in fault detection of nonlinear systems.
【学位授予单位】：南京师范大学
【学位级别】：硕士
【学位授予年份】：2011
【分类号】：TH165.3

【引证文献】