Non-parametric modeling of industrial systems using variable discrimination in wavelet

• Main Author: Cabigon, Noel Papas
• Format: text
• Language: English
• Published: Animo Repository 2001
• Subjects: Wavelets (Mathematics); Mathematical models; Systems engineering; Chemical Engineering
• Online Access: https://animorepository.dlsu.edu.ph/etd_doctoral/942
• Institution: De La Salle University

This paper presents a new non-parametric modeling technique. The method is simple and yet efficient and robust compared to existing non-parametric modeling procedures such as artificial neural network (ANN), principal component regression (PCR), and the traditional stepwise regression. It can solve regression problems with multiple collinearities and improve models by removing redundant parameters that other methods cannot handle. The procedure is: first, discriminate the variables to separate the collinear and non-collinear variables by examining the behavior of the variance inflation factor (VIF) or the diagonal of the correlation matrix in the rectified wavelet coefficients in scale. Collinear variables' VIF increase in scale while those noncollinear variables show fluctuating or slightly decreasing trend. The best model subsets are created from the results of discrimination. Then the best model is selected from the best model subsets using the criteria relative mean square error (rmse), PRESS residual, PRESS, Fstat, R2adj and the number of explanatory parameters included in the model. Application of the procedure is illustrated using computer-generated data with up to 20 and up to three explanatory and response variables, respectively. Other tests made include data from published researches of model with bilinear, quadratic and integer terms. Two industrial studies are also presented. These are modeling of the secondary coating line for fiber optic cable production of Eupen Cable Asia, Incorporated and the occurrence of high chloride in the National Power Corporation Tiwi Geothermal Power Plant. The method worked successfully. The variability in the type of systems were the methods works could mean that the method can be extended to other types of system regardless of the nature of the parameters included.