Predictive modeling the internal bond of medium density fiberboard using a modified principal component analysis

• Published in: Forest products journal 2008-04, Vol.58 (4), p.49-55
• Main Authors: Clapp, N.E. Jr, Young, T.M, Guess, F.M
• Format: Article
• Language: English
• Subjects: algorithms; correlation; Correlation analysis; Cost control; Cost reduction; Data processing; Eigenvalues; Eigenvectors; equations; Forest products industry; Influence; internal bond strength; Laboratories; Lumber industry; Manufacturers; Manufacturing; mechanical properties; medium density fiberboards; prediction; Predictive control; principal component analysis; Principal components analysis; process variables; product quality; Raw materials; Real time; Regression analysis; Relational data bases; Sensors; statistical models; strength (mechanics); Studies; testing; thickness; Variables; Wood; Wood composites
• ISSN: 0015-7473; 2376-9637

In this paper, real-time process data are aligned in time-order with destructive test data to reduce cost by better predictive modeling. A modified principal component analysis (PCA) is used to develop an empirical model to predict the internal bond of MDF based on a selected subset of process variables. These process variables are selected by picking variables with the highest absolute correlations with internal bond. Our modified PCA is used on these selected standardized process variables to obtain transformed composite variables or modes. The 10 modes are reduced to three using correlation criteria and the three best modes are used to generate an empirical model to predict internal bond. Results for the most produced thickness category of MDF are presented primarily, while some comments are made on two other thickness categories. The root mean square error relative to the mean of each category varied from 9.3 percent to 11.2 percent, which are quite helpful improvements in this manufacturing setting. More attention to the collection of the current process variables via information quality efforts might be useful for additional future improvements. Even though the plant had 179 quantitative process variables, our PCA and correlation analyses suggests some other variable(s) need exploring and collecting to further reduce these error rates. [PUBLICATION ABSTRACT]