Predicting effects of selected impregnation processes on the observed bending strength of wood, with use of data mining models

Abstract

Wood materials have been used in many products such as furniture, stairs, windows, and doors for centuries. There are differences in methods used to adapt wood to ambient conditions. Impregnation is a widely used method of wood preservation. In terms of efficiency, it is critical to optimize the parameters for impregnation. Data mining techniques reduce most of the cost and operational challenges with accurate prediction in the wood industry. In this study, three data-mining algorithms were applied to predict bending strength in impregnated wood materials (Pinus sylvestris L. and Millettia laurentii). Models were created from real experimental data to examine the relationship between bending strength, diffusion time, vacuum duration, and wood type, based on decision trees (DT), random forest (RF), and Gaussian process (GP) algorithms. The highest bending strength was achieved with wenge (Millettia laurentii) wood in 10 bar vacuum and the diffusion condition during 25 min. The results showed that all algorithms are suitable for predicting bending strength. The goodness of fit for the testing phase was determined as 0.994, 0.986, and 0.989 in the DT, RF, and GP algorithms, respectively. Moreover, the importance of attributes was determined in the algorithms.

Full Article

Predicting Effects of Selected Impregnation Processes on the Observed Bending Strength of Wood, with Use of Data Mining Models

Selahattin Bardak,^a Timucin Bardak,^b,* Hüseyin Peker,^c Eser Sözen,^d and Yıldız Çabuk ^d

Wood materials have been used in many products such as furniture, stairs, windows, and doors for centuries. There are differences in methods used to adapt wood to ambient conditions. Impregnation is a widely used method of wood preservation. In terms of efficiency, it is critical to optimize the parameters for impregnation. Data mining techniques reduce most of the cost and operational challenges with accurate prediction in the wood industry. In this study, three data-mining algorithms were applied to predict bending strength in impregnated wood materials (Pinus sylvestris L. and Millettia laurentii). Models were created from real experimental data to examine the relationship between bending strength, diffusion time, vacuum duration, and wood type, based on decision trees (DT), random forest (RF), and Gaussian process (GP) algorithms. The highest bending strength was achieved with wenge (Millettia laurentii) wood in 10 bar vacuum and the diffusion condition during 25 min. The results showed that all algorithms are suitable for predicting bending strength. The goodness of fit for the testing phase was determined as 0.994, 0.986, and 0.989 in the DT, RF, and GP algorithms, respectively. Moreover, the importance of attributes was determined in the algorithms.

Keywords: Wood material; Bending strength; Mechanical properties; Data mining; Optimization

Contact information: a: Sinop University, Faculty of Engineering and Architecture, Department of Computer Engineering, 57000, Sinop, Turkey; b: Bartin University, Bartin Vocational School, Furniture and Decoration Program, 74200, Bartin, Turkey; c: Artvin Çoruh University, Department of Forest Industry Engineering, Faculty of Forestry, 08000, Turkey; d: Bartin University, Faculty of Forestry, Department of Forest Industrial Engineering, 74200, Bartin, Turkey; *Corresponding author: timucinb@bartin.edu.tr

INTRODUCTION

The chemical modification of wood has received much attention since the mid-twentieth century. Usually, chemical modification in solid wood is made for dimensional stability and biological resistance. Despite many studies in the literature, there have been only limited studies of industrial applications (Gérardin 2016). This is mainly due to the application difficulties of wood preservation processes such as acylation and the emergence of environmental, economic, and technical problems in the transition from laboratory to industry. Factors such as wood type, impregnation agent, impregnation time, temperature, vacuum, diffusion, solvent, concentration ratio, and retention ratio affect the mechanical properties of wood (Rowell 2009). Simsek et al. (2010) carried out mechanical and decay tests of beech wood (Fagus orientalis L.) and Scots pine wood (Pinus sylvestris L.) treated with environmentally friendly boron compounds. They reported that all the different concentrations of boron compounds applied decreased the bending strength compared with the control sample. In a similar study, Adanur et al. (2017) stated that different proportions of boron compounds decrease the bending resistance and increase the screw holding strength. In another study, Tan et al. (2017) examined static bending resistance and dynamic bending resistance properties of pine and beech wood using 1%, 3%, and 5% barite. The static bending strength values of pine and beech woods increased by 55% and 83% in 3% and 5% concentrations compared with control samples, respectively.

Technological developments necessitate interdisciplinary interaction. In the light of the available information, making new decisions and making predictions are important. Many complex real-world problems have been solved by data mining (Predic et al. 2018; Zhang et al. 2019). It is an automated analysis of data sets to expose relationships that are both understandable and useful (Hand and Mannila 2001; Wei and Watkins 2011). Data mining tasks can be classified as description or prediction. The purpose of the prediction is to find a model to estimate the values of future events (Rémy et al. 2018). Data mining techniques are often more powerful, flexible, and effective than statistical techniques for information discovery (Kantardzic 2011). Decision tree (DT), random forest (RF), and Gaussian process (GP) algorithms are widely used for predicting. These algorithms are easy to interpret and fast to calculate (Höppner et al. 2020). In data science, the DT is defined as a classification procedure. This classification algorithm recursively partitions a data set into smaller subdivisions based on a set of tests defined at each branch in the tree (Pu et al. 2018). RF is an ensemble method in machine learning that involves construction (growing) of multiple decision trees via bootstrap aggregation (Shaikhina et al. 2019). The RF algorithm is based on decision trees and combined with aggregation and bootstrap ideas. The RF algorithm maintains low bias on the training dataset by creating a collection of unpruned decision trees (Nadi and Moradi 2019). GP based on statistical learning theorem is a machine learning method. It is mainly used to calculate the covariance between the data points used in the model. It is suitable for high dimensional complex regression problems (Zhang et al. 2019).

Data analysis techniques provide useful information in wood science. It is important to understand how production components affect each other to solve problems in the wood industry. Very little research has been done on the prediction of the mechanical performance of wood materials. Tiryaki and Hamzacebi (2014) studied the bending strength of the heat-treated wood as predicted by artificial neural networks (ANNs), which successfully predicted bending resistance. Atoyebi et al. (2018) used ANNs to examine the physical and mechanical properties of particleboards and the impact of various factors on production. The study showed that ANNs have great potential in predicting the mechanical properties of particleboards.

In this paper, the RF, GP, and DT models were created to estimate the bending strength of impregnated wood. The relationships between wood type, vacuum time, and diffusion time were examined with these models.

EXPERIMENTAL

Material and Method

Material

Imported Scots pine and wenge logs were used. First, the logs were cut into battens. The cut slats were cut radially, and all samples were obtained from sapwood. Barite (BaSO₄) was obtained from Gülmer Mining (Bilecik, Turkey) in powder form. Scots pine is native and coniferous wenge is foreign and leafy tree species. The reason for choosing these tree species was to be able to compare the leafy and coniferous and native and foreign tree species.

Preparation of Impregnation Solution

The impregnation solution was prepared with 1% barite on a heated magnetic stirrer. The solution was prepared at 220 °C for 30 min and allowed to stand at room temperature (25 °C) for 24 h.

Impregnation Method

Impregnation was carried out as described in ASTM D 1413 (1976). Before the impregnation, all samples were coded and weighed with a 0.01 mm precision analytical balance. The samples were oven-dried at 103 ± 2 °C. Fully dry samples were impregnated with four different vacuum times (10, 20, 30, 40 min) and pressurization times (diffusion) (25, 35, 45, 55 min) in the impregnation boiler. Since vacuum and diffusion time are important variables that affect the impregnation of wood species, these two parameters were especially chosen. Another reason for choosing these parameters is to determine the effect of different vacuum and diffusion times on the bending strength. Impregnation conditions were carried out according to Taghiyari et al. (2013); 600 mm-Hg vacuum was applied as a pressure of 0.6 MPa. A total of 320 samples were prepared, including two tree species (2), different vacuum times (4), diffusion times (4), and 10 samples for each variation (2 x 4 x4 x5 = 160).

Bending Strength

Wood test samples and bending tests were prepared according to TS 2470 (1976) and TS 2474 (1976), respectively. Five samples were prepared from each variation from two different tree species, four different vacuum times, four different diffusion times, for a total of 160 (5 x2 x4 xx4) samples subjected to static bending resistance.

Data Collection

The experimental data used in this study were the measurements obtained by the bending tests of the wood material impregnated under different conditions. The data set consisted of 160 records. There were three attributes (wood type, vacuum time, and diffusion time) that feature in mechanical property prediction and one attribute serves as the output (bending strength). Table 1 contains a summary of the values of the numeric attributes from the training data set.

Models

Three algorithm models were selected to prediction bending strength of impregnated wood: decision trees (DT), random forest (RF), and Gaussian process (GP). These models were based on establishing the relationship between independent and dependent variables using training methods. In this study, 70% of all data were used for training and 30% for testing purposes. All models were developed with RapidMiner Studio Version 9.3 software (Boston, USA), which has been used in many studies (Phark et al. 2018; Cuesta et al. 2019). RapidMiner Studio consists of operators and each operator has a task. Operators are added end-to-end to prepare the process workflow. Several parameters have to be set when using algorithms as predictive modelling in this software. To find the best parameters, the optimize parameters (Grid) operator was used. Thus, the best parameters were determined for each model separately as shown in Fig. 1; the optimal parameters for all models are listed in Table 2.

Fig. 1. The process workflow used to optimize the parameters of the models

After determining the optimal parameters, the process was created to compare the models. Figure 2 shows the process workflow used to compare models.

Fig. 2. The process workflow used to compare models

Simulations were created based on the models. The aim was to find the input values for the highest bending strength. Rapidminer software allows for preparing real-time simulations. Figure 3 shows the process workflow prepared for simulation.

Fig. 3. The process workflow prepared for simulation

Finally, weights of attributes (wood type, vacuum time, and diffusion time) were determined with DT and RF algorithms. Thus, the most important feature for the prediction was found.

Model Evaluation

Goodness of fit (R²), root mean square error (RMSE), and mean square error (MSE) were used to evaluate the estimation accuracy of each model, as follows,

(1)

(2)

MSE = RMSE² (3)

where Y_O and Y_P are the measured and predicted values, respectively, and the bar denotes the mean of the variable.

RESULTS AND DISCUSSION

Mechanical Properties

Wood type, vacuum time, and diffusion time affected the bending strength of impregnated wood material. Table 3 presents the bending strength depending on changes in the wood type, vacuum time, and diffusion time. Increasing vacuum and diffusion times caused decreases in bending strength of both Scots pine and wenge wood. Under the same conditions, the bending strength of wenge wood gave higher values than Scots pine.

Results of the Models

Three machine learning models were established using the experimental data. All models were trained and tested with the same data sets, and the predictive performance of models were compared. The estimated results of the three data mining models tested for wenge and Scots pine wood are given in Tables 4 and 5, respectively. For wenge wood, percentages of correct predictions were found as 99.55%, 99.37% and 99.38% for decision tree, random forests, and Gaussian process models, respectively. For Scots pine wood, percentages of correct predictions were found as 99.49%, 99.38%, and 99.46% for decision tree, random forests, and Gaussian process models, respectively.

Three different assessment criteria were used to evaluate all estimation models (DT, RF, GP). In the classification problem, these performance measures are widely used (Caballero et al. 2017; Shafaei et al. 2019). Various performance measures related to the DT, RF, and GS models are shown in Table 6.

MSE and RMSE are measures of error each. Therefore, low results are measures showing high performance in inverse proportion to performance (Wang and Xu 2004; Gultepe 2019).

After determining the performance of the models, the simulation process was performed separately. Figure 4 shows the simulation screen for the decision tree model. The term simulation differs from the term modeling. Simulation can be defined as the representation of a process. The simulation process is done to achieve three goals. First, users better understand complex models such as deep learning. Second, users check whether the model is behaving as expected. Third, users find the most appropriate input settings to achieve the desired result.

Fig. 4. Screenshot of simulation of decision tree model

Determination of Optimum Conditions

The values of the attributes for the highest bending strength were determined with all models. Table 7 shows the attributes and values for the highest bending strength.

All models determined similar input attributes (Diffusion time, Vacuum Time, and Wood Type) for the highest bending strength. Simulation results can be used to support decision-making in the impregnation process. The best parameters are determined according to the needs of the enterprises so that limited resources can be used more effectively.

The weights of attributes were determined by decision trees and random forest algorithms. attributes and weight values, where each weight represents the feature importance for the given attribute (diffusion time, vacuum time, and wood type). Table 8 shows the weight of attributes for Decision Trees and Random Forest algorithms.

Discussion

According to the bending strength tests performed in the study, the bending strength of wenge wood was 15% higher than that of Scots pine. In wenge and Scots pine woods, increasing the vacuum time from 10 min to 40 min resulted in a gradual decrease in bending strength values. Bending strength of Scots pine woods applied to vacuum for 40 min was 108.53 N/mm², and when the vacuum was reduced to 10 min, bending resistance increased by 10.2% to 119.68 N/mm². Under the same conditions, this increase in wenge wood was determined as 8.81%. The effect of diffusion time applied on bending strength showed the same effect in both tree species. Reducing the diffusion time from 55 min to 25 min resulted in a 2.6% increase in bending strength of the two wood species. In wenge wood species, the highest bending strength (140.83 N/mm²) was obtained in 10 min vacuum and 25 min diffusion conditions. The lowest bending resistance value was obtained under 40 min vacuum and 55 min diffusion conditions. The highest (120.65 N/mm²) and the lowest (106.90 N/mm²) bending strength values of Scots pine wood were obtained under the same conditions where the highest and lowest values obtained in the wenge wood. As a result, increasing diffusion and vacuum times decreases the bending strength values of the tree species used in the study, which was consistent with previous reports. Aydemir et al. (2016) impregnated Scots pine, ash and Iroko wood with boron compounds and reported that impregnated wood exhibited higher strength properties (MOR) than control samples.

According to the modelling performance results, the highest estimation accuracy was seen in DT (Table 6). Also, other models showed similar results to DT. All models are suitable for estimating the bending strength of impregnated wood material. In the literature, an R² value greater than 0.9 represents a very satisfactory model (Leachtenauer et al. 1997; Heng and Suetsugi 2013). The success of the models provides evidence of the benefits of data mining in the wood impregnation industry. The most successful model parameters are shown among the models considering that the MSE value approaches 0 (Gultepe 2019). When the MSE values are examined, it is seen that the model that gives the values closest to 0 is the decision tree. In addition, multiple regression model was created. In this way, the prediction performance of statistical analysis (multiple regression) and the data mining models (decision tree random forests and Gaussian process) were compared. The prediction performance of the multiple regression model (R² = 0.985, RMSE = 1.042, MSE = 1.086 for testing phase) was found to be somewhat lower than the decision tree model (R² = 0.994, RMSE = 0.680, MSE = 0.462 for testing phase), random forests model (R² = 0.986, RMSE = 1.028, MSE = 1.057 for testing phase) and Gaussian process model (R² = 0.989, RMSE = 0.926, MSE = 0.857 for testing phase).

According to the results obtained from the algorithms (Table 8), the most important factors are diffusion time, vacuum time, and wood species, respectively. Multi-attribute problems must be solved to identify the weights of attributes. Zavadskas et al. (2010) reported that data mining algorithms can be used for this purpose.

Future research should examine the performance of these data mining algorithms in the prediction of bending strength of impregnated wood material in more complex conditions. The anisotropic structure of wood offers sufficient options to create different conditions. Sapwood-heartwood ratio, early wood-late wood ratio, annual ring characteristics, density, and moisture content are some of them.

CONCLUSIONS

1. Experiments were performed by varying three impregnation parameters: the wood type, the diffusion time, and the vacuum time. The highest average bending strength obtained was 140.83 N/mm² at the 10 min vacuum time and 25 min diffusion time with wenge wood.
2. Bending strength of impregnated wood material was successfully estimated by data mining techniques. The use of data mining algorithms in the impregnation of wood can greatly increase productivity because prediction algorithms respond to the best inputs for each situation.
3. Three different prediction models (DT, RF, GP) were compared according to R², MSE, and RMSE performance measurements. The highest success in the estimations was observed in DT algorithm with 0.994 R² value for the testing phase. It was concluded that prediction algorithms can affect the optimization of the impregnation process positively.
4. The importance of the factors was defined as diffusion time, vacuum time, and wood type vacuum duration, respectively.

REFERENCES CITED

Adanur, H., Fidan, M. S., and Yaşar, Ş. Ş. (2017). “The technological properties of oriental beech (Fagus orientalis Lipsky) impregnated with boron compounds and natural materials,” BioResources 12(1), 1647-1661. DOI: 10.15376/biores.12.1.1647-1661

ASTM-D 1413-99 (1976). “Standard test method of testing wood preservatives by laboratory soil block cultures, ” ASTM International, West Conshohocken, USA.

Atoyebi, O. D., Awolusi, T. F., and Davies, I. E. (2018). “Artificial neural network evaluation of cement-bonded particle board produced from red iron wood (Lophira alata) sawdust and palm kernel shell residues,” Case Studies in Construction Materials 9, e00185 (In press). DOI: 10.1016/j.cscm.2018.e00185

Aydemir, D., Civi, B., Alsan, M., Can, A., Sivrikaya, H., Gunduz, G., and Wang, A. (2016). “Mechanical, morphological and thermal properties of nano-boron nitride treated wood materials,” Maderas. Ciencia y tecnología 18(1), 19-32. DOI: 10.4067/S0718-221X2016005000003

Caballero, D., Pérez-Palacios, T., Caro, A., Amigo, J. M., Dahl, A. B., ErsbØll, B. K., and Antequera, T. (2017). “Prediction of pork quality parameters by applying fractals and data mining on MRI,” Food Research International 99, 739-747. DOI: 10.1016/j.foodres.2017.06.048

Cuesta, H. A., Coffman, D. L., Branas, C., and Murphy, H. M. (2019). “Using decision trees to understand the influence of individual-and neighborhood-level factors on urban diabetes and asthma,” Health & Place 58, 102119. DOI: 10.1016/j.healthplace.2019.04.009

Gérardin, P. (2016). “New alternatives for wood preservation based on thermal and chemical modification of wood—A review,” Annals of Forest Science 73(3), 559-570. DOI: 10.1007/s13595-015-0531-4

Gultepe, Y. (2019). “A comparative assessment on air pollution estimation by machine learning algorithms,” European Journal of Science and Technology 16, 8-15. DOI: 10.31590/ejosat.530347

Hand, D. J. Mannıla, H., and Smyth, P. (2001). Principles of Data Mining (Adaptive Computation and Machine Learning), MIT Press, Cambridge, MA, USA.

Heng, S., and Suetsugi, T. (2013). “Using artificial neural network to estimate sediment load in ungauged catchments of the Tonle Sap River Basin, Cambodia,” Journal of Water Resource and Protection 5(2), 111-123. DOI: 10.4236/jwarp.2013.52013

Höppner, S., Stripling, E., Baesens, B., and Verdonck, T. (2020). “Profit driven decision trees for churn prediction,” European Journal of Operational Research 284(3), 920-933. DOI: 10.1016/j.ejor.2018.11.072

Kantardzic, M. (2011). Data Mining: Concepts, Models, Methods, and Algorithms, John Wiley & Sons, Hoboken, NJ, USA.

Leachtenauer, J. C., Malila, W., Irvine, J., Colburn, L., and Salvaggio, N. (1997). “General image-quality equation: GIQE,” Applied Optics 36(32), 8322-8328. DOI: 10.1364/AO.36.008322

Nadi, A., and Moradi, H. (2019). “Increasing the views and reducing the depth in random forest,” Expert Systems with Applications 138, 112801. DOI: 10.1016/j.eswa.2019.07.018

Phark, C., Kim, W., Yoon, Y. S., Shin, G., and Jung, S. (2018). “Prediction of issuance of emergency evacuation orders for chemical accidents using machine learning algorithm,” Journal of Loss Prevention in the Process Industries 56, 162-169. DOI: 10.1016/j.jlp.2018.08.021

Predic, B., Ilic, M., Spalevic, P., Trajkovic, S., Jovic, S., and Stanic, A. (2018). “Data mining based tool for early prediction of possible fruit pathogen infection,” Computers and Electronics in Agriculture 154, 314-319. DOI: 10.1016/j.compag.2018.09.023

Pu, Y., Apel, D. B., and Lingga, B. (2018). “Rockburst prediction in kimberlite using decision tree with incomplete data,” Journal of Sustainable Mining 17(3), 158-165. DOI: 10.1016/j.jsm.2018.07.004

Rémy, N. M., Martial, T. T., and Clémentin, T. D. (2018). “The prediction of good physicians for prospective diagnosis using data mining,” Informatics in Medicine Unlocked 12, 120-127. DOI: 10.1016/j.imu.2018.07.005

Rowell, R. M. (ed.) (2009). Handbook of Wood Chemistry and Wood Composites, Taylor & Francis, London.

Shafaei, S. M., Loghavi, M., Kamgar, S. (2019). “Feasibility of implementation of intelligent simulation configurations based on data mining methodologies for prediction of tractor wheel slip,” Information Processing in Agriculture 6(2), 183-199. DOI: 10.1016/j.inpa.2018.10.004

Shaikhina, T., Lowe, D., Daga, S., Briggs, D., Higgins, R., and Khovanova, N. (2019). “Decision tree and random forest models for outcome prediction in antibody incompatible kidney transplantation,” Biomedical Signal Processing and Control 52, 456-462. DOI: 10.1016/j.bspc.2017.01.012

Simsek, H., Baysal, E., and Peker, H. (2010). “Some mechanical properties and decay resistance of wood impregnated with environmentally-friendly borates,” Construction and Building Materials 24(11), 2279-2284. DOI: 10.1016/j.conbuildmat.2010.04.028

Taghiyari, H. R., Enayati, A., and Gholamiyan, H. (2013). “Effects of nano-silver impregnation on brittleness, physical and mechanical properties of heat-treated hardwoods,” Wood Science and Technology 47(3), 467-480. DOI: 10.1007/s00226-012-0506-7

Tan, H., Ulusoy, H., and Peker, H. (2017). “The effects of impregnation with barite (BaSO 4) on the physical and mechanical properties of wood materials,” Journal of Bartin Faculty of Forestry 19(2), 160-165. DOI: 10.24011/barofd.346900

Tiryaki, S., and Hamzacebi, C. (2014). “Predicting modulus of rupture (MOR) and modulus of elasticity (MOE) of heat-treated woods by artificial neural networks,” Measurement, 49, 266-274. DOI: 10.1016/j.measurement.2013.12.004

TS 2470 (1976). “Wood-sampling methods and general requirements for physical and mechanical tests,” Turkish Standards Institution, Ankara, Turkey.

TS 2474 (1976). “Wood – Determination of ultimate strength in static bending,” Turkish Standards Institution, Ankara, Turkey.

Wang, W., and Xu, Z. (2004). “A heuristic training for support vector regression,” Neurocomputing 61, 259-275. DOI: 10.1016/j.neucom.2003.11.012

Wei, W., and Watkins Jr, D. W. (2011). “Data mining methods for hydroclimatic forecasting,” Advances in Water Resources 34(11), 1390-1400. DOI: 10.1016/j.advwatres.2011.08.001

Zavadskas, E. K., Turskis, Z., Ustinovichius, L., and Shevchenko, G. (2010). “Attributes weights determining peculiarities in multiple attribute decision making methods,” Engineering Economics 66(1), 32-43. DOI:

Zhang, M., Li, M., Shen, Y., Ren, Q., and Zhang, J. (2019). “Multiple mechanical properties prediction of hydraulic concrete in the form of combined damming by experimental data mining,” Construction and Building Materials 207, 661-671. DOI: 10.1016/j.conbuildmat.2019.02.169

Article submitted: March 2, 2021; Peer review completed: May 1, 2021; Revised version received and accepted: May 7, 2021; Published: May 14, 2021.

DOI: 10.15376/biores.16.3.4891-4904