Real-time process modeling of oriented strand board panels
Andre, N. O., and T. M. Young.  2014.  PTF BPI 2014. 3rd Intern. Conf. Processing Tech. for the Forest and Bio-based Products Industries, Salzburg/Kuchl, Austria, 24-26 September 2014. p. 393-400.

Abstract:
This study focuses on the real-time prediction of parallel elasticity index (EI) and edge swell for an oriented strand board (OSB) manufacturing process. 11mm sheathing and 18mm flooring products were modeled. As wood composite panel plants periodically test their products, a data fusion application was developed to properly align laboratory mechanical test results and their corresponding process data. Fused data are employed to build regression models that yield real-time predicted mechanical and physical properties values when new process data are available. The modeling algorithm core uses genetic algorithm or correlation to the quality attribute to preselect a meaningful subset of process variables. Multiple cali-bration models are then built using several regression methods. A comparison of process variables selection methods showed that correlation to the property of interest was yielding the best regression model results. Among four available regression methods, neural networks consistently returned high correlation coefficients and low prediction errors, edging all the other methods. Recent edge swell calibration model for 11mm sheathing product returned a 0.94 correlation coefficient, an average prediction error of 1.33%, and a mean normalized average error of 4.85%. Parallel EI calibration model for 11mm sheathing product returned a 0.88 correlation coefficient, and average prediction error of 26Nm2/m, and mean normalized average error of 3.52%. Edge swell calibration model for 18mm flooring product returned a 0.94 correlation coefficient, an average prediction error of 0.41%, and a mean normalized average error of 3.35%. Parallel EI calibration model for 18mm flooring product returned a 0.93 correlation coefficient, an average prediction error of 66.13Nm2/m, and a mean normalized average error of 1.93%.