Prototype of the Near-Infrared Spectroscopy Expert System for Particleboard Identification

'e overall goal of this work was to develop a prototype expert system assisting quality control and traceability of particleboard panels on the production floor. Four different types of particleboards manufactured at the laboratory scale and in industrial plants were evaluated. 'e material differed in terms of panel type, composition, and adhesive system. NIR spectroscopy was employed as a pioneer tool for the development of a two-level expert system suitable for classification and traceability of investigated samples. A portable, commercially available NIR spectrometer was used for nondestructive measurements of particleboard panels. Twenty-five batches of particleboards, each containing at least three independent replicas, was used for the original system development and assessment of its performance. Four alternative chemometric methods (PLS-DA, kNN, SIMCA, and SVM) were used for spectroscopic data classification.'emodels were developed for panel recognition at two levels differing in terms of their generality. In the first stage, four among twenty-four tested combinations resulted in 100% correct classification. Discrimination precision with PLS-DA and SVMC was high (>99%), even without any spectra preprocessing. SNV preprocessed spectra and SVMC algorithm were used at the second stage for panel batch classification. Panels manufactured by two producers were 100% correctly classified, industrial panels produced by different manufacturing plants were classified with 98.9% success, and the experimental panels manufactured in the laboratory were classified with 63.7% success. Implementation of NIR spectroscopy for wood-based product traceability and quality control may have a great impact due to the high versatility of the production and wide range of particleboards utilization.


Introduction
Particleboard is a panel product manufactured from lignocellulosic materials, combined with an adhesive system and bonded together under heat and pressure.Particleboards are easy to process and flexible in application.ey are mainly used for furniture production, but in combination with other materials, they might be used for parquet, insulation materials, sheathing boards for timber framed walls, packaging, or "do-it-yourself products."e major types of particles used to manufacture particleboard include wood shavings, flakes, wafers, chips, and sawdust [1].
A new strategy of EU aims at increasing the share of bioenergy in the EU's total energy production.It is expected that 20% of the energy produced in Europe after 2020 will be from renewable resources, where 80% of the above will be related to lignocellulosic feedstock.Simultaneously, a significant increase in the production volumes for other woodbased products (pulp, boards, and furniture) is expected.erefore, even if supply and demand of wood on the market is balanced at present, it will most probably not be in equilibrium within the coming years [2].As a consequence, several alternative resources, mainly agricultural and industrial residues, fast-growing shrubs and plantations, and postconsumer wood, have received considerable attention in recent years [3].Effective use of bagasse [4], oil palm waste [5], bark [6] paper sludge, [3], kenaf stalks [7] wheat straw [8], needle litter [9], vine pruning's [10] waste tissue paper, and corn peel [11] among others was previously reported.Such materials are ecological, functional, and environmental-friendly; therefore, they serve as sustainable raw materials for manufacturing particleboards.e type of raw resources used for panels manufacturing, its quality, size of particles, moisture content as well as type and amount of bonding system have significant effect on particleboard properties [12].
e most important quality assessment aspects of manufactured panels are emissions, which mainly depends on the type and amount of resin [1].e recent trend to reduce the formaldehyde release from manufactured wood products has led to the substitution of ureaformaldehyde (UF) resin with several alternatives [12].Liquefied wood was previously reported as an interesting replacement for commonly used adhesives [13][14][15].e influence of liquefied wood on particleboard properties was previously investigated by the authors [16,17].However, alternative quality control tools with potential application for online process control are still desired.
Portable spectroscopic equipment operating in the NIR range is a highly interesting technology for the wood-based sector [18].Meder et al. [19] reported successful implementation of FT-NIR spectroscopy for at-line measurement for quality control of melamine-urea-formaldehyde resin in composite wood-panel production.Taylor and Via [20] used visible and near-infrared spectroscopy to quantify phenol formaldehyde resin content in oriented strandboard.Campos et al. [21] used FT-NIR to evaluate composition of agro-based particleboards.Janiszewska et al. [22] reported the application of FT-NIR spectroscopy for clustering raw materials used for liquefaction and their transformation products.Even though several examples of NIR use for product and process quality monitoring have been reported, the majority of them rely on FT-NIR instruments that are relatively costly (both time and investment wise) and with limited applicability for inline measurement.
Numerous chemometric techniques for multivariate classification and discrimination were reported as suitable for processing of NIR spectra.ese are usually divided into discriminant or class-modelling methods.In discriminant analysis, an unknown sample is always assigned and can be allocated only to one of the classes given in the training set.e class-modelling approach is more flexible.e sample can be accepted by more than one class model and therefore be recognized as confused.e most common classification methods are linear discriminant analysis (LDA), partial least-squares discriminant analysis (PLS-DA), k-nearest neighbours (kNN), SIMCA, and SVM [23].
LDA is a probabilistic method, which assumes that each sample belonging to a particular class follows a multivariate Gaussian distribution.It requires the explicit calculation of this probability for the formulation of the classification rule.
e main disadvantage of this method is the fact that it requires a significantly higher number of training samples than the number of variables, the variables themselves cannot be correlated, and all categories have the same within-class scatter, which is hard to assure in many experimental cases [23].
PLS-DA is a multivariate inverse least-squares analysis method used to classify samples.It decomposes the spectra as linear combinations of principal components (PCs) that express the majority of the information (variability) contained in the global dataset.e predictor variables or latent variables (LVs) are generated from the input variables to maximize the variance between sample classes in the model [24].
KNN is nonparametric and instance-based algorithm.is means that it does not make any assumption on the underlying data distribution and does not use the training data to do generalization.e kNN makes decision based on the entire training data set.
e object is classified by a majority vote of neighbours.According to Adeniyi et al. [25], the main advantages of this method are capability of handling training data that are too large to fit in memory, use of simple Euclidean distance to measure the similarities between training and test data, providing a faster and more accurate recommendation as a result of straightforward application.
Soft independent modelling of class analogy (SIMCA) is the most common supervised modelling method, representing class-modelling approach.It requires a training data set of samples with a set of attributes and their class membership.In SIMCA, a principal component analysis (PCA) is performed on each class in the data set; thus, a principal component model is used to represent each class.In SIMCA, there is no restriction on the number of measurement variables, and few samples per class are enough to run the model.Data classification is a common task in machine learning.
e support vector machine (SVM) is a very flexible method that makes no assumption regarding data.It is a nonlinear classification method that constructs a set of hyperplanes in a high-or infinite-dimensional space.Good separation is achieved by the hyperplane that has the largest distance to the nearest training data point of any class [24].It works by obtaining the optimal boundary of two groups in a vector space independent of the probabilistic arrangements of vectors in training set.When the linear boundary in lowdimension input space is not enough to separate two classes, SVM can create a hyperplane that allows linear separation in the higher dimension feature space [26].SVM has been successfully used for data mining, pattern recognition, and artificial intelligence fields [18].
Practical implementation of discrimination methods in the wood-based industry is a challenging task.Wood is a heterogeneous and anisotropic biological material.e correct implementation of NIR spectroscopy for bio-based materials requires a complex approach [27].A two-level expert system is proposed here for identification of particleboard panels with a portable, commercially available NIR spectrometer.e overall goal of this work was to develop a prototype system that might assist quality control and traceability of particleboard panels on the production floor.

Materials and Methods
Four different types of particleboards manufactured in the laboratory and in industrial plants were evaluated.e materials differed in terms of panel type (single-layer or three-layer), composition (industrial particles, recycled wood, or alternative lignocellulosic materials including fastgrowing species), and the adhesive system (ureaformaldehyde (UF) resin, UF resin modified with liquefied wood (LW), or UF resin modified with starch).A summary of investigated panels is presented in Table 1.In total, 170 samples representing 25 variants were investigated.A colour image of representative samples surface for each variant is presented in Figure 1.

Particleboards Manufactured with Recycled Wood (D).
Industrial wood particles obtained from a local woodprocessing sawmill, containing 25% recycled wood content, were used as a raw material for particleboard production.
e particles were sorted using an Allgaier vibration screening machine with screens of mesh diameters 8, 2, 1, and 0.5 mm.Particle fractions ≤8 mm and ≥1 mm were selected for particleboard production.
e adhesive was prepared as a mixture of an industrial urea-formaldehyde resin and liquefied wood (10-20% relative to the dry weight of the resin).e liquefied wood was prepared from four types of wood-processing industry wastes: mixed hardwood-softwood powder (LWP), pine (LP), beech sawdust (LB), and bark (LB) [16].
e industrial urea-formaldehyde glue resin characteristics were as follows: gel time of 75 seconds, viscosity 336 mPa•s, total solids content 69.4%, and pH 7.3.Ureaammonium nitrate solution (46%) was used as a curing agent, constituting 1% of the resin dry mass.
Single-layered particleboards of 12 mm thickness with 0, 10, 15, and 20% liquefied wood content in the adhesive resin were produced in the laboratory.e nominal density of the panels was 650 kg•m −3 .All boards were conditioned after pressing at 20 °C and 65% relative humidity.Nine types of single-layer panels were prepared with different adhesives system, as listed in Table 1.Each panel type was produced in two independent batches.Six replicates for each panel were analysed; in total, 108 samples were investigated.

Particleboards Manufactured with Fast-Growing
Wood Species (S).Single-and three-layer particleboard panels were manufactured from Eastern red cedar (Juniperus virginiana L.) using 9% urea-formaldehyde, or a combination of 15% modified corn starch and 2% urea-formaldehyde adhesive.ree types of samples were prepared: type A: single-layer board with 9% UF; type B: three-layer board with 9% UF; and type C: three-layer board with a combination of 15% starch and 2% UF.All boards were produced in a laboratory press using a pressure of 5 MPa, at a temperature of 165 °C for 5 and 10 minutes in the case of modified corn starch bonded samples [28].Six independent replicates for each panel were analysed with an NIR spectrometer.
e stalks of lignocellulosic materials were reduced in size with a Pallmann's chipper.Fractions smaller than 10 mm and bigger than 1 mm were used for the core of the panels.e chips were used for manufacturing four types of three-layer boards of 16 mm thickness, with the raw density of 680 kg•m −3 [29].Six independent replicates for each panel were analysed with an NIR spectrometer.

Industrially Manufactured Particleboards (P).
ree-layer urea-formaldehyde resin-bonded particleboards, type P2, suitable for non-load-bearing purposes in dry areas, manufactured by six diverse manufacturing plants of a corporation were used as reference industrial samples.In addition, panels with different thicknesses (38 mm, 28 mm, 18 mm, and 8 mm) prepared by a single producer were investigated.ree replicates for each panel type were analysed using NIR spectroscopy.

Spectroscopic Measurement and Data Mining.
e MicroNIR 1700 compact sensor produced by Viavi Solutions (Santa Monica, CA, USA) was used for spectroscopic measurements.Each spectrum was measured as an average of 10 consecutive scans.
e scanning frequency was 50 Hz, corresponding to 20 ms of integration time.
e spectral range was from 950 to 1650 nm (10526-6060 cm −1 ).128 spectral points were defined for each spectrum and corresponded to pixels of the CCD detector of the instrument.Ten independent measurements were done on each panel assuring measurements at different locations over its surface.Each set of 10 spectra was preprocessed with extended multiplicative scatter correction (EMSC) and after that averaged to homogenize the spectral fingerprint of heterogeneous surfaces of panel.
e MicroNIR instrument has the proven potential for infield and inline applications due to its rigid construction and integration of all optical, electronic, and mechanical components.
PLS_Toolbox 8.0 (Eigenvector Research, Manson, WA, USA) and LabView 13 (National Instruments, Austin TX, USA) were used for data processing and mining.Four discriminant analysis methods: PLS-DA, SVM, KNN, and SIMCA, were used for spectra classification.Models were calculated considering use of the raw spectra and spectra preprocessed with normalization, standard normal variate (SNV), EMSC, and 1 st and 2 nd derivatives (Savitzky-Golay algorithm, 2 nd polynomial order, and 15 smoothing points).e order of data was randomized and then divided into calibration (66%) and independent validation (34%) data sets.e samples used for model validation were all different from those used for calibration.

General Concept of the Expert System Architecture.
Expert systems are innovative process tools that enhance the user's productivity in accomplishing a task or solving a problem.ese vary in at least two important dimensions, including knowledge and technological complexities [30].Knowledge complexity is mainly determined by the degree of depth and specialization of the internalized knowledge, the scope of the decision, and the level of expertise required to solve the problem.e technological complexity includes diversity of hardware and software, the complexity of the user's environment, the scale of the software design effort, the complexity of required database accesses, and special user interfaces among others [30].
e main goal of this work was to develop an expert system capable of detecting and tracing different particleboards.e general concept of such a system implemented in the furniture factory is presented in Figure 2.
e system relies on the spectroscopic measurements acquired during arrival of the new batch to the factory floor with the easy-to-handle and portable instrument.e particular characteristics of the instrument investigated within this project allow its direct implementation in the real-life applications.It is recommended to average several NIR spectra collected from the characterized sample in order to minimize the heterogeneity effect of the complex surfaces of particleboards.A spectrum acquired in that way is considered as a fingerprint of the distinct batch and is recorded in the database for further analysis.e usage of spectra is twofold.Firstly, it is applied for the further improvement of the chemometric models by providing new case data.Secondly, it is used for future tracking and identification of elements during production as well as in final products.
e NIR spectrum of an unknown sample is confronted at the first level of the expert system with broad classes representing different particleboard manufacturers or panel types.In the case of successful identi cation, the same spectrum is an input for the second-level chemometric model, where a nity to the speci c batches recorded previously is assessed.Such a numerical tool might be highly useful for the quality control of panels arriving to the production oor, as well as for authentication of their origin and composition.e demonstrative implementation of the proposed expert system was performed within this research project.e statistical evaluation of the discriminations success rate, performed at both system levels, is presented in the following section.

Prototype Expert System for Particleboard Identi cation.
e concept of the abovementioned expert system implemented for demonstration of its feasibility is presented in Figure 3.It consists of ve chemometric models implemented at two discrimination levels.Model #1 is used to screen the unknown spectrum and determine the most probable class corresponding to the particleboard type.Four classes were therefore de ned at the rst level of the expert system corresponding to "sample sources" as summarized in the rst column of Table 1.e second and more speci c discrimination is executed afterward by implementing one of the second-level models (#2 to #5).
e number of classes in each of these models corresponds to the number of previously de ned batches.It has to be mentioned that it is critical to continuously feed the expert system with the most recent spectra of new batches in order to assure constant improvement of the discrimination system and proper identi cation of all historically used particleboard batches.
An expert system approach to quality control at the industrial scale was proposed by Paladini [31].He described three-step quality control approach that included precontrol activities, inspection, and a decision stage.Liukkonen et al. [32] described the intelligent optimization and modelling system for electronics production.
ey proposed three modules consisting of appropriate mathematical tools speci cally tailored to each task: preprocessing, variable selection, and optimization.A data-driven approach was proposed to achieve proactive quality improvement of the production process.Hobballah et al. [33] proposed a casual map to show what variables should be included in the design optimization and how the components interact causally.is approach was successfully used for the preliminary design of an insulating composite mat based on wood bers.

Discriminant Analysis.
Four alternative classi cation methods: KNN, PLS-DA, SVM, and SIMCA, were tested for the optimal NIR spectral data classi cation.All these methods are supervised techniques, as they use prede ned information about the class membership for all samples selected for model calibration.In that way, the model can be tuned to classify new unknown samples in one of the known classes on the basis of its individual pattern [26].Discriminating techniques usually build models based on all the categories concerned for the discrimination.
erefore, samples can be classi ed into one of the prede ned categories, even if actually they do not belong to any of these [26].It is possible, however, to de ne a certain threshold for a nity to any of the classes.In that case, samples not passing the threshold may be classi ed as unde ned.
Particleboard panel discrimination at the producer level (model #1) di erentiated classes in terms of manufacturing conditions (laboratory or industrial scale), di erent resources used for panels manufacturing (industrial particles, addition of recycled wood, fast-growing wood species, agricultural wastes, etc.), or di erent adhesives used for panel production.
e discrimination success scores between di erent panel providers are summarized in Table 2, where selected spectra preprocessing and diverse discrimination algorithms are presented.Four among twenty-four tested combinations resulted in 100% correct classi cation.e optimal combinations were PLS-DA and SIMCA with 2 nd derivative spectra, and SVM with both SNV and EMSC preprocessing.Furthermore, discrimination precision with PLS-DA and SVM was high (>99%), even without any spectra preprocessing.
e second level of the expert system was designed to identify speci c batches.Twenty-ve lots of particleboards were modelled by four independent chemometric models (#2 to #5), developed for each panel provider separately.
e SVM algorithm classifying SNV preprocessed NIR spectra was implemented.
e summary of the model's performance is presented in Figure 4, where confusion tables indicate the ability of each model to properly classify (or misclassify) experimented samples from the validation set.All samples of panels manufactured by producer S and G were correctly classi ed.Industrial panels produced by di erent manufacturing plants of the producer P were classi ed with 98.9% success and the panels manufactured by producer D with 63.7% success.It is important to mention that, in the case of the producer D, the same substrate (industrial particles containing 25% of recycled wood) was used for manufacturing all investigated panels.e only di erence was the adhesive system, with 0, 10, 15, and 20% share of lique ed wood addition in the UF resin.

Implementation of the Prototype Expert System.
e implementation of the expert system in real conditions is a challenging task.Qian et al. [34] proposed a four-step process that consists of the following: (i) Knowledge representation (including all variables characterizing production process) (ii) Database development (including all decision rules, functions, and metadata) (iii) Machine interface design (in uencing the quality of the expert system for fault diagnosis and the realtime response of the system) (iv) Knowledge maintenance (including update, verication, and correction of the errors in knowledge base) e experiences gained when developing chemometric models discriminating diverse particleboards stimulated extension of this research to a prototype expert system ready for testing in real-world applications.e hardware used for the prototype (portable NIR spectrometer and computer) was similar to the laboratory tests.e di erences were in the software, where the original code was customised.It included integration of tools for the control of the sensor setup, data acquisition, spectra processing, and discriminant analysis.e required chemometric models (#1 to #5) were computed o ine with PLS_toolbox and were exported as Matlab scripts to LabView.e algorithm of the prototype expert system for the batch identi cation is presented in Figure 5. e software allows acquisition of the new NIR Journal of Spectroscopy spectra for feeding the database as well as assessment of the unknown spectrum in one of the prede ned classes.

Conclusions
e traceability of particleboards is an important practical issue a ecting economic performance of furniture producers and related industries.It allows control of material quality and ensures high standards of nal products.NIR spectroscopy was employed here as a pioneer tool for development of the expert system suitable for assisting such a traceability.Twenty-ve batches of particleboards, each containing at least three independent replicas, were used for the original system development and for its performance assessment.
e hardware and software developed was working properly, and it was implemented as a ready-to-use prototype system.A two-stage expert system seems to be optimal for implementation of the particleboard discrimination as it allows a rough (in the rst stage) and detailed (in the second stage) classi cation of particleboard samples to the level of a single batch.In two cases (two producers), 100% correct classi cation was achieved at the second stage.
It has to be mentioned that a drawback of the system presented is the necessity to condition samples before measurement.It includes both refreshing of the surface by gentle sanding and conditioning of samples at wellde ned climatic conditions.It is due to the fact that NIR spectroscopy is extremely sensitive to the water molecules presented in the measured samples.Nevertheless,

Figure 2 :
Figure 2: Concept for the traceability of particleboards in the furniture industry.

Figure 3 :
Figure 3: Prototype expert system's layout for the experimental set of particleboard samples.

Figure 4 :
Figure 4: Confusion table of the particleboard batch prediction with the prototype NIR expert system: project LIDER-D (a), OSU-S (b), ITD-G (c), and industry-P (d).

Table 1 :
Manufactured panels investigated in this research.

Table 2 :
e results of the discrimination success rate at the rst level of expert system.