Prediction of amino resin solids content with PLS based on NIR: Improving model performance using a data balancing strategy

Abstract

Monitoring the solids content (SC) of amino resins plays a significant role in reducing costs and increasing production efficiency in the wood-based panels (WBP) industry. The goal of this study was to use NIR spectroscopy and regression-based methods for predicting SC in amino resins. Several wavenumber intervals were investigated to determine the best regions of the spectrum for model improvement. To address dataset imbalances, an oversampling technique was used, resulting in a more accurate representation of underrepresented SC values in the initial dataset.

The calibration and test set splitting were performed using the random (RD) method, as well as the Sample set Partitioning based on joint X–Y distances (SPXY) and Kennard-Stone (KS) methods, which improved model reliability by providing calibration data that covered the entire input space. Predictive models were developed using Partial Least Squares (PLS) regression, with the number of latent variables optimized through 10-fold cross-validation. Combining wavenumber interval selection, oversampling, and the KS split enabled significantly improved prediction performance metrics and robustness. This approach provides an effective alternative for the WBP industry, allowing for more efficient and robust quality control of amino resins. The best model had a maximum absolute error of 0.3 % on the test set, which was comparable to the performance of the reference method, demonstrating its potential for use in industrial applications.