The aim of this study was to propose a nonlinear multiscale principal component analysis (NLMSPCA) methodology for process monitoring and fault detection based upon multilevel wavelet decomposition and nonlinear principal component analysis via an input-training neural network. Prior to assessing the capabilities of the monitoring scheme on a nonlinear industrial process, the data is first pre-processed to remove heavy noise and significant spikes through wavelet thresholding. The thresholded wavelet coefficients are used to reconstruct the thresholded details and approximations. The significant details and approximations are used as the inputs for the linear and nonlinear PCA algorithms in order to construct detail and approximation conformance models. At the same time non-thresholded details and approximations are reconstructed and combined which are used in a similar way as that of the thresholded details and approximations to construct a combined conformance model to take account of noise and outliers. Performance monitoring charts with non-parametric control limits are then applied to identify the occurrence of non-conforming operation prior to interrogating differential contribution plots to help identify the potential source of the fault. A novel summary display is used to present the information contained in bivariate graphs in order to facilitate global visualization. Positive results were achieved.
Dissertation (M Eng (Control Engineering))--University of Pretoria, 2007.