Real-time predictive modeling of industrial engineered wood quality parameters using genetic algorithms and multivariate statistical methods
Andre, N. O., and T. M. Young.  2009.  63rd Forest Products Society International Meeting. June 23. Boise, ID.

This paper presents data mining-based multivariate calibration methods for predicting quality parameters from industrial engineered wood process variables. An essential prerequisite for successful modeling resides in the proper fusion of a real-time process variables database and a laboratory test results database using transact SQL code and appropriate time lagging. Genetic algorithms paired with partial least squares regression dictate the selection of important process variables based on several optimization criteria. Neural networks and several other multivariate methods are employed to build calibration models between the selected process variables and the corresponding quality response variable. Real-time predictions are compared with the corresponding laboratory test results, automatically triggering new optimization iteration if relative errors between predicted and actual values are too high. Several case studies using real plant data will be presented and discussed.