Modeling of Compressive Strength Parallel to Grain of Heat Treated Scotch Pine ( Pinus sylvestris L . ) Wood by Using Artificial Neural Network

In this study, the compressive strength of heat treated Scotch Pine was modeled using artifi cial neural network. The compressive strength (CS) value parallel to grain was determined after exposing the wood to heat treatment at temperature of 130, 145, 160, 175, 190 and 205oC for 3, 6, 9, 12 hours. The experimental data was evaluated by usi ng multiple variance analysis. Secondly, the effect of heat treatment on the CS of samples was modeled by using artifi cial neural network (ANN).


INTRODUCTION 1. UVOD
Heat treatment is a wood modifi cation method used to improve some properties of the wood.Heat treatment also helps to diminish equilibrium moisture content of wood samples (Mazela et al., 2004).The temperature level and duration of heat treatment mostly change from 180 to 280 °C and from 15 min to 24 h depending on the heat treatment process, sample size, wood species, moisture content of the sample and the properties of the fi nal product (Kandem et al., 2002;Militz et al., 2002).
The values of hardness and strength of wood decrease with the increase of heat treatment parameters (temperature and duration).These effects are achieved especially when heat treatment is carried out for a long time.The strength values of wood most affected by heat treatment are impact and static bending strengths, while the least affected property is the modulus of elasticity (Korkut et al., 2008).
Artifi cial neural network (ANN) is a computational model based on the information processing system of the human brain.ANN model is composed of three layers, which are called input layer, hidden layer and output layer.This network structure is also called MLP (Işeri and Karlık, 2009).
While the input layer receives the initial values of the variables, the output layer shows the results from the network for the input.The hidden layer carries out the operation design to achieve the output.The number of neurons in the input layer must correspond to the number of entry variables, and the output layer must have as many neurons as the number of outputs manufactured by the network.However, there is no rule to allow prior decisions to indicate the number of neurons contained in the hidden layer or sublayer.The only way to obtain the hidden layer is by a process of trial and error (Sha, 2007).
Artifi cial neural network has been widely used in many wood industries, such as in the wood identifi cation system (Tou et al., 2007;Khalid et al., 2008;Estaben et al., 2009a;Juni or et al., 2006) in the suggestion on the application of geodesy (Arslan et al., 2007), in the prediction of wood dielectric loss factor (Avramidis et al., 2006), in the calculation of wood thermal conductivity (Xu et al., 2007), in predicting fracture toughness of wood (Samarasinghe et al., 2007), in the evaluation of strength of wood timbers (Tanaka et al., 1996), in the prediction of bending strength and stiffness in western hemlock (Shawn et al., 2007), in the prediction of particle-board mechanical properties (Fernández et al., 2008), in the optimization of process parameter in a particleboard manufacturing process (Cook et al., 2000), in the detection of structural damage in medium density fi berboard panels (Long et al., 2008), in the prediction of modulus of rupture and modulus of elasticity of fl ake board (Yapıcı et al., 2009).It has also been applied to obtain the hygroscopic equilibrium points (Avramidis and Iliadis, 2005), to classify wood defects (Drake and Packianather, 1998), to determine the internal bond values of particleboard (Cook and Chiu, 1997;Fernandez et al., 2008), and in statistical process control in the manufacture of particleboard (Estaben et al., 2009b).
In this study, compression strength parallel to grain of heat treated Scotch pine wood samples was examined experimentally, and then artifi cial neural network (ANN) system was designed for predicting this value.

MATERIJALI I METODE
Scotch pine wood (Pinus sylvestris L.) was chosen randomly from timber merchants of Karabuk, Turkey.In the selection of wood material, special emphasis was on the properties of non-defi cient, proper, knotless, normally grown wood (without zone line, reaction wood, decay, insect and mushroom damages).The selected specimens were cut to sizes of 20×20×300 mm and they were exposed to heat treatment at 130, 145, 160, 175, 190 and 205 ºC for 3, 6, 9, and 12 hours.Then, they were resized to 20×20×30 mm.The compressive strength values were determined from test samples according to TS 2595 standard (TS 2595).

Statistical analyses 2.1. Statističke analize
Data for each test were statistically analyzed.Analysis of variance was used to test the signifi cance between factors and levels.When the analysis of variance pointed a signifi cant difference among the factors and levels, a comparison of the means was conducted employing a Tukey test.

Design of artifi cial neural network for CS value 2.2. Dizajn umjetne neuronske mreže za tlačnu čvrstoću
In this study, the effects of heat treatment conditions on compressive strength parallel to grain of scotch pine wood were determined experimentally.Secondly, artifi cial neural network model was applied The aim of the network is to predict CS of test samples.The network is trained by using MATLAB neural network module (nftool).A total of 70 % of these data is used for training, 15 % is used for validation, 15 % is used for testing.The data for each class are chosen randomlly from the total data set.
In this study, the number of neurons in the hidden layer is 25.This number is obtained by trial and error.In the nftool nature hyperbolic tansig function f (x) =1/ (1+exp(-x)) is applied.Input data is applied after normalization process between -1 and +1.

REZULTATI I RASPRAVA
The air dry density of Scots pine is 0.62 g/cm 3 .The CS values of Scotch pine woods, obtained from experimental results, were compared with ANN model to determine the accuracy of the developed model.Our network was trained with designed data set to obtain the predicted result.Regression analysis of the training phase is given in Figure 2.
It is seen that regression coeffi cients obtained from training, validation and test phase of network are calculated close to 1.This result showed that the designed model is reliable.
Based on this comparison, the developed model agreed with average test results at the accuracy level of 97.02 % of CS value.Both experimental values and prediction values are given in Table 1.
The variance analysis of CS based on heat treatment circumstances was done by using variance analysis (Table 2.).The difference between the groups regarding the effect of variance sources on CS was signifi cant ( =5 %).
It can be seen that the conditions of heat treatment has no effects on the CS values of Scotch pine wood according to variance analysis.So, the results of the Tukey test conducted to determine the importance of the differences between the groups are given in Table 3.
It can be seen that the CS values ranged between 45.46 N/mm 2 and 52.29 N/mm 2 according to Tukey's test (Table 3).It can be stated that when the temperature and time of the heat treatment increase, the value of CS increases.However, the ratio of increase is not statically signifi cant.Thus, they are put into the same homogeneous group.The change of CS values, both experimental and prediction values, are given in Figure 3.

ZAKLJUČAK
Based on the results of tests, it can be said that the properties of compression strength parallel to grain were slightly affected by applying the heat treatment.It can be seen that the CS values decrease with the increase of the time and temperature of heat treatment.The values obtained from experimental work are used for artifi cial neural network system.CS values of test samples have been predicted by the designed model at 97.33 % accuracy level.So, ANN model can be used to predict many mechanical and physical properties of wood and wood composite materials.