| Process monitoring is required both to detect process abnormal changes as early as possible and to reduce the number of false alarms. Multivariate statistical process control (MSPC) methods have been widely applied to satisfy these key demands. Among them, PCA is the most popular one, which relates to its conceptual simplicity. However, when applying PCA to industrial process monitoring, a lot of problems appear because of its performance limit. Therefore, many scholars have proposed some improved methods, such as kernel PCA, recursive PCA and consensus PCA which extend applications of PCA in the industry process control. This dissertation develops the research based on the predecessor’s work. The main research contents are as follows:(1) The process statistical model built by conventional KPCA is time-invariant, while most real industrial processes are slowly time-varying. When a static KPCA model is used to monitor processes with the dynamic changes, false-alarms often occur, which significantly reduce the reliability of the method. To overcome the false-alarms, a recursive KPCA (RKPCA) algorithm is proposed based on the recursive singular value decomposition (SVD) technique in the kernel space. The RKPCA is suitable for online model updating by computing recursively the eigenvalues and eigenvectors of covariance matrix for the training data in the feature space. The simulation result of the process monitoring shows this method can not only reduce the false-alarms, but also improve the effect of the fault detection.(2) As large-scale industrial processes get more and more complex, the problem is that the result can be hard to interpret owing to the large number of variables. Consensus kernel PCA (CKPCA) can improve the interpretability and reduce the complexity of multivariate models by monitoring processes in a hierarchical manner. However, because of the effect of the time-varying process condition in real industrial processes, CKPCA is not available for dynamic processes. Thus false-alarms may occur. To solve this problem, an approach is proposed which first utilizes CKPCA to build the process statistical model and then uses RKPCA to update the CKPCA model recursively by combining CKPCA and RKPCA. At last, this algorithm is applied to fault detection and has a good effect which proves the method i(?) effective and feasible. |
