Prediction of internal bond strength in a medium density fiberboard process using multivariate statistical methods and variable selection

Abstract

This paper presents new data mining-based multivariate calibration models for predicting internal bond strength from medium density fiberboard (MDF) process variables. It utilizes genetic algorithms (GA) based variable selection combined with several calibration methods. By adopting a proper variable selection scheme, the prediction performance can be improved because of the exclusion of non-informative variable(s). A case study using real plant data showed that the calibration models based on the process variables selected by GA produced better performance than those without variable selection, with the exception of the radial basis function (RBF) neural networks model. In particular, the calibration model based on supervised probabilistic principal component analysis (SPPCA) yielded better performance (only when using GA) than partial least squares (PLS), orthogonal-PLS (O-PLS), and radial basis function neural networks models. The SPPCA model benefits most from the use of GA-based variable selection in this case study.