Study On Modified Kernel Independent Component Analysis Based Non-Gaussian Process Monitoring Method

With the expansion and the complication of modern industrial, and the development of senor technology and storage technology, the monitored data of industry process is becoming more and more abundant. The data-based process monitoring methods have been greatly accelerated the development along the last decade. Among the methods, the Multivariate Statistical Process Control (MSPC) methods have been widely applied to solve the problem of process monitoring. But the non-Gaussianity of modern industry process data often limited the using of traditional MSPC method which under the assumption of the samples obeyed Gaussian distribution. Independent component analysis (ICA) method has significant advantage on the handling of non-Gaussian data. It is promising to apply ICA to industry process monitoring. The higher requirement of process monitoring, and the nonlinear, time varying, multimode and other features of industry process variables, require further improvement of original ICA to apply it to process monitoring. This dissertation develops the research based on the predecessor’s work, according to the non-Gaussian process monitoring method’s problems on fault isolation and multimode process monitoring, the main research contents are listed as follows:(1) Based on the foundation of traditional fault reconstruction direction, this article proposed a modified KICA based fault isolation method, which extract fault relevant directions from independent component subspace, realized the fault isolation of new fault data. Using the character of rich information in independent component, the history normal and fault data are mapped into independent component subspace. Through analyzing the relationship between fault ICs and normal ICs, the fault relevant ICs is found, and used as fault feature direction. The proposed method is applied to fault isolation of two kinds of fault data in electrical fuse magnesia furnace. Finally, compared with traditional contribution method and KPCA based fault reconstruction method, the proposed method has better fault isolation capability.(2) Aiming at the problem of multimode process monitoring problem in non-Gaussian process, this article proposed a multimode kernel independent component analysis (MKICA), which takes the advantage of KICA in handling non-Gaussian data, and the character of multimode process. Through comparing the difference to monitoring statistic along different modes, the independent component subspace and residual subspace are separated into common information part and different information part. And these separated parts are modeled respectively to monitor. The proposed method is applied to fault detection of three modes in Tennessee Eastman Process, and compared to global scale modeling method. The test result of proposed method shows that, the proposed method is better than global scale modeling method, and the common model has better university to each mode, while the different model of each mode has higher sensitivity to fault detection.