White Rose University Consortium logo
University of Leeds logo University of Sheffield logo York University logo

Feature extraction and knowledge discovery in process operation analysis

Chen, Bing Hui (1998) Feature extraction and knowledge discovery in process operation analysis. PhD thesis, University of Leeds.

[img] Text
uk_bl_ethos_506283.pdf

Download (13Mb)

Abstract

An integrated framework for process monitoring is developed in this study which consists of three components: (1) feature extraction from dynamic transient signals using multiscale wavelet transform; (2) operational state identification using unsupervised and recursive learning methods; and (3) automatic extraction of knowledge rules from process operational data and embedding of the extracted knowledge in the structure and weights of fuzzy-neural networks. The methodologies and the prototype system which have been developed are illustrated and evaluated using data collected from a dynamic simulator of a refinery catalytic cracking process. Methods for pre-processing dynamic transient signals for feature extraction, dimension reduction and noise removal are investigated and a new method is developed which makes use of wavelet transform to determine the singularities and irregularities of a dynamic transient signal by identifying the extrema from wavelet multiresolution analysis. The method is able to reduce the dimensionality of the data and removes noise components in a single step as well as capturing the most significant components of the dynamic response. A modified version of the unsupervised neural network ART2, designated ARTNET, has been developed which uses wavelet feature extraction to provide a substitution of the data pre-processing part of ART2. ARTNET is shown to be more effective in avoiding the adverse effects of noise, less sensitive to user defined parameters and faster in computation, as well as still retaining the advantages of unsupervised and recursive learning. Based on this, a fuzzy neural network is developed which is able to automatically extract knowledge rules from process data. The knowledge rules which are generated are transparent and explicit to operators. The method is therefore able to bridge the gap between numerical data and qualitative knowledge and takes advantage of the features of neural networks for capturing concepts and so provides an effective and robust method for learning knowledge from process data. Various methods for integrating different facets of a problem, and making use of this information in parallel to mutually compensate for drawbacks of any single approach are also exploited. Data obtained from a dynamic simulator of a refinery fluid catalytic cracking process (FCC) has been used to illustrate the methodologies and to evaluate a prototype system for using these new approaches. FCC provides a very useful case study because of the highly non-linear dynamics arising from the strong interactions between the reactor and fluidised bed regenerator derived from the mass and momentum balance. The use of simulation data makes it possible to look at the results in detail so that the methods can be fully tested. The case studies illustrate the potential of the methods developed.

Item Type: Thesis (PhD)
Academic Units: The University of Leeds > Faculty of Engineering (Leeds) > School of Process, Environmental and Materials Engineering (Leeds)
Depositing User: Ethos Import
Date Deposited: 02 Mar 2010 15:44
Last Modified: 07 Mar 2014 11:21
URI: http://etheses.whiterose.ac.uk/id/eprint/626

Actions (repository staff only: login required)