Problem description Pulp and paper are forest products made from wood fibres, which go through a series of chemical and mechanical treatments in order to reach their final state. The forest products industry largely manufactures pulp and paper as intermediate products, which are supplied to other manufacturing industries for further refinement into various end use products. Some typical end use products are various types of plastics, rayon, pharmaceuticals, food additives, and finished papers (from newsprint to high quality paper). To serve such a broad range of applications a typical mill usually produces pulp and paper at a variety of grades, with each grade having its own end use properties.
Adhering to the unique specifications laid-out by the end use product manufacturers is of utmost importance. As a result, pulp and paper manufacturers perform various tests at every stage of the process to ensure that the exact specifications are met. Furthermore, to gain customer satisfaction a large number of complex quality control tests are also performed on the finished products.
The majority of these quality control tests are preformed in analytical laboratories, where complicated wet-chemistry procedures are employed to determine the product's chemical properties. Some of these tests are time-consuming and require constant attention. Such stringent demands can induce undesirable experimental errors, which reflect on the perceived quality of the finished product. Most wet-chemistry tests are naturally destructive to the solid samples since they require dissolving the sample into alkaline or acidic solutions. As a result, multiple samples are required to obtain a complete chemical analysis of the finished product.
To overcome the above-mentioned issues with wet-chemistry procedures the forest products industry is aggressively exploring new technologies for quality control of finished pulp and paper samples. Ideally such technology should enable rapid testing of the samples, and simultaneously provide multiple properties from a single sample in a non-destructive manner. NIR imaging techniques are a very promising solution to this problem.
Classification of paper samples A set of specially created paper samples obtained from ASTM (The American Society for Testing Materials) was used in this study. These samples were originally created for ASTM's research program to study the aging of printing and writing papers. The composition of each sample was varied in a designed experiment fashion, where 9 different combinations of 6 ingredients have been used to create the paper dataset and each combination was used to produce several reams of 8.5 by 11 paper sheets. The objective is to investigate the ability of NIR imaging spectroscopy to extract pre-defined composition information from paper samples, and use it to classify the samples based on differences in their chemical make-up.
The NIR spectrum used for the study was 933 nm to 1663 nm, discretized into 110 individual wavelength images. The approximate surface area imaged was 155 mm (L) 110 mm (W), represented by 500 pixels (L) and 126 pixels (W). Three images were acquired for each of the paper samples. Figure 1 illustrates an example of a multi-spectral NIR image at five of the 110 NIR wavelengths. The NIR spectra of five selected pixels (as highlighted on Figure 1(a)) are illustrated in figure 2.
Figure 1: An example multi-spectral NIR reflectance image acquired at 5 wavelengths
Figure 2: NIR reflectance spectra of 5 selected pixels from paper sample 1 (spatial locations of pixels as marked in Figure 1(a)). Vertical red lines represent the wavelengths at which the images in figure 2 are sampled.
It was decided to use the first MPCA loading vector as the feature vector for classification of the paper samples. This vector represents overall chemical information across the imaged section of the paper. Figure 3 shows that the 9 sample groups, representing unique ingredient compositions, are well separated into their respective point clusters. Furthermore, all the samples can be seen to follow the orientation defined by lines A, B, and C in figure 3. Compared with the chemical content of each samples, It was found out that the direction from line A to B (i.e. along line C) explains increasing paper brightness due to their Hardwood BCTMP pulp content, whereas paper samples along line B contain Bleached Hardwood Kraft pulp.
Figure 3: t1-t2 score space scatter plot of PCA to classify feature vectors of 27 paper samples
Prediction of pulp properties Four end properties of finished pulp were selected for prediction: S10, S18, DCM Resin, and Intrinsic Viscosity. S10 and S18 measure the alkaline solubility of finished pulp and are indicators of degraded cellulose and hemi-cellulose remaining in the pulp. The analytical laboratory testing procedure to measure S10 and S18 requires dissolving the dry pulp into alkaline solutions and performing various measurements on the solutions to determine the amount of degraded cellulose and hemi-cellulose. Obtaining the final S10 and S18 measurements for each pulp sample requires approximately 1.5 hours of experimental procedures.
DCM Resin is a measure of the amount of remaining resinous materials in finished pulp, which are extractable with organic solvents like Dichloromethane. The analytical laboratory testing procedure for measuring DCM Resin involves dissolving the solid pulp sample into the organic solvent and extracting the solution via siphoning at regular intervals. The final DCM Resin measurement for each pulp sample requires approximately 4 hours of experimental procedures.
Intrinsic Viscosity measures the average molecular chain length of the polymers making up the pulp fibres. The solid pulp sample is torn and dissolved in various reagents prior to measurement of intrinsic viscosity. The experimental procedure to obtain a measurement typically requires 20 minutes.
A total of 60 pulp samples were imaged over the three days of production during an at-line experimental run. The samples belonged to one of two main pulp grades: (1) Rayon Grade, or (2) Pharmaceutical Grade. Additionally, the rayon grade was further divided into two sub-grades having different pulp end property specifications. The grade change between the two main grades was also captured during the at-line experimental run. A surface area of 140 mm (L) by 93 mm (W) was captured as 448 pixels (L) by 102 pixels (W) NIR reflectance images in 108 unique wavelengths spanning the 933 - 1650 nm range.
A strategy for predicting laboratory tested pulp properties from multi-spectral NIR pulp images is shown in figure 4. First, features are extracted from image data, and then the extracted feature variables are regressed against properties using partial least squares (PLS).
Figure 4: Schematic of strategy for predicting laboratory tested pulp properties (Y) from multi-spectral NIR pulp images (X)
The prediction results for one of the properties, S18 are shown in figure 5. It can be seen that the model performs reasonably well in following the pulp property trends from Grade 1A to Grade 1B to Grade 2.
Figure 5: Time series plot of S18 pulp property laboratory measured data and 2OSC+PLS1(2PC) regression model
Related publications Manish Bharati, John F. MacGregor and Marc Champagne (2002). Using near-infrared multivariate image regression techniques to predict pulp properties, submitted to
TAPPI Journal, September 2002.
Manish Bharati, John F. MacGregor, Marc Champagne and M. Barrete (2002). Using NIR multivariate image regression techniques to predict pulp properties. Presented at
Control Systems 2002, Stockholm, Sweden, 3-5 June 2002.
