Fault Diagnosis And Quality Monitoring Based On Multivariate Statistical Analysis

Fault diagnosis and quality monitoring are the important technical support to ensure product quality and safety of industrial production process. Compared with the traditional fault diagnosis methods which depend on mathematical model or expertise, data-driven multivariate statistical pro-cess monitoring technology has been causing more and more engineering and academic attentions and has gradually become the major requirement and research focus in fault diagnosis because it just depends on normal offline or online data, takes the more sophisticated multivariate statisti-cal analysis methods as the main measure and suits to industrial processes with highly correlated multivariate.Despite the multivariate statistical analysis methods have been used to fault diagnosis for a long time, a lot of work are all around the problem how to break through limitations, and improve existing methods to better suit the actual processes; while some nature problems have not been thoroughly studied, such as the importance and detection sensitivities of different variables are dif-ferent, the effect of principal component analysis (PCA) subspace decomposition to fault diagnosis and the effectiveness of partial least squares (PLS) subspace decomposition. Based on previous s-tudies, this article in-depth studies several fundamental questions of multivariate statistical process monitoring technology, and proposes a variety of effective fault detection, fault diagnosis and qual-ity monitoring methods for different applications:1. In terms of the contribution degree of process variables to quality variables, the affect ability of variables on T2/Q detection index or PCA statistical model, process variables are decomposed in detail, and a design method for weighted factors in relative PCA (RPCA) is presented. The pro-posed method increases the relative importance and relative sensitivity of the important insensitive variables and decreases the relative importance and relative sensitivity of the unimportant sensitive variables. It can help to improve the fault detection effect of key variables and thereby improve the quality prediction performance.2. The effect of PCA subspace decomposition on the fault detection capability has been addressed and a subspace decomposition principle based on eigenvalues and eigenvectors elements has been established. The residual subspace has been decomposed in detail and a multi-space detection method based on PCA has been presented. It improves the significance of failure in residual subspace and then improves the detection effect of minor fault. Based on historical normal data, a multi-space fault diagnosis method based on PCA by using different process variables contribution degree to latent variables has been further proposed, which can realize both fault detection and fault diagnosis to a certain extent.3. For the batch process with time-series characteristic, a modal division and process mon-itoring method based on information gain is proposed, which compensates the inadequate of tra-ditional modal division methods effectively and decreases the limitations that sample obey nor-mal distribution and modal unchanged of multi-PCA modal method. For the multi-modal process with modal disorder and frequent switching, a real-time union fault detection method is present-ed, which makes up the shortage of tailored multi-modal method which difficult to online identify modal and then difficult to switch monitoring model and adopt appropriate methods to monitor and further improves the multi-modal process real-time monitoring methods.4. Aiming at the problem of quality-related process monitoring, on the basis of the analysis of the effectiveness of PLS subspace decomposition, two directly latent space projection methods are established. According to the main changes of quality variables are caused by process variables, on the basis of perform PCA on quality variables, quality latent space are generated and then guide process variables decomposition. Aiming at the problem that parts of quality changes are caused by process variables, quality prediction value by using system identification method is given, and then PCA is performed on quality prediction value and the nonlinear process monitoring is promoted. The new method not only have simple model but only can obtain better effect with less cost.5. As for the problem of fault detection of a class of continuous-time non-uniform sam-pling system, by combining subspace identification and equivalent space technique, a direct design residual generator method only based on input and output data is proposed and fault detection is applied, which can realize rapid residual generation and reduce corresponding equivalent matrix dimensions at the same time. This method is no need to identify system model. It decreases the calculation and simplifies design process at the same time. Furthermore, it is an useful exploration and promotion for data- and model-based fault diagnosis technologies.Numerical simulations and TEP case or penicillin fermentation process are uses in this thesis to verify the effectiveness of the proposed methods. At last, the main contents of this article is concluded and several challenging research topics are proposed in the author’s opinion.