Artificial neural network and partial least square regressions for rapid estimation of cellulose pulp dryness based on near infrared spectroscopic data.

The content of water in fiber suspension and affects pulp refining, bleaching and draining operations. Cellulose pulp dryness estimate through near infrared (NIR) spectroscopy coupled with multivariate regressions or artificial neural network (ANN) techniques are not well explored yet. In this study models were developed to estimate cellulose pulp dryness in pads based on the NIR spectra. Thus, the cellulose pulp pads (4 mm thick) were weighed and their NIR spectra were obtained in several stages during desorption from 13.1 to 98.3% of content of solids. Partial least square regression (PLS-R) was developed from whole NIR spectra (1300 Absorbance values) and six spectral variables (from 1300) were selected for developing the PLS-R (6) and the ANN model. Both trained neural network and regression can predict pulp dryness of unknown cellulose pulp pads from their NIR data with an error of 2.5%. PLS-R models based on whole NIR spectra showed accurate predictions (the R² of lab-determined and estimated values plot was 0.99) while the ANN showed the same predictive performance from only six NIR variables. Predictive models developed from full NIR spectra and those based on only 6 variables were compared. Our findings indicate that NIR spectroscopy coupled with multivariate analysis and Artificial neural networks are a promising tool for monitoring the weight variation due to dewatering of the cellulose pulps in real time.