Abstract
This study aimed to predict the CNC cutting conditions for the best wood surface quality, energy, and time savings using artificial neural network (ANN) models. In the CNC process, walnut, and ash wood were used as materials, while three different cutting tool diameters (3 mm, 6 mm, and 8 mm), spindle speed (12000 rpm, 15000 rpm, and 18000 rpm), and feed rate (3 m/min, 6 m/min, and 9 m/min) were determined as cutting conditions. After the cutting processes were completed with the CNC machine, energy consumption and processing time were determined for all groups. Surface roughness and wettability tests were performed on the processed wood samples, and their surface qualities were determined. The experimentally obtained data were analysed in ANN, and the models with the best performance were obtained. By using these prediction models, optimum cutting conditions were determined. Using the findings of the study, the optimum cutting condition values can be determined for walnut and ash wood with the smoothest and best wettable surface. Furthermore, in CNC processes using such materials, minimum energy consumption and shorter processing time can be obtained with optimum cutting conditions.
Download PDF
Full Article
Prediction of Optimum CNC Cutting Conditions Using Artificial Neural Network Models for the Best Wood Surface Quality, Low Energy Consumption, and Time Savings
Evren Osman Çakıroğlu,a Aydın Demir,b İsmail Aydın,b and Ümit Büyüksarı c,*
This study aimed to predict the CNC cutting conditions for the best wood surface quality, energy, and time savings using artificial neural network (ANN) models. In the CNC process, walnut, and ash wood were used as materials, while three different cutting tool diameters (3 mm, 6 mm, and 8 mm), spindle speed (12000 rpm, 15000 rpm, and 18000 rpm), and feed rate (3 m/min, 6 m/min, and 9 m/min) were determined as cutting conditions. After the cutting processes were completed with the CNC machine, energy consumption and processing time were determined for all groups. Surface roughness and wettability tests were performed on the processed wood samples, and their surface qualities were determined. The experimentally obtained data were analysed in ANN, and the models with the best performance were obtained. By using these prediction models, optimum cutting conditions were determined. Using the findings of the study, the optimum cutting condition values can be determined for walnut and ash wood with the smoothest and best wettable surface. Furthermore, in CNC processes using such materials, minimum energy consumption and shorter processing time can be obtained with optimum cutting conditions.
DOI: 10.15376/biores.17.2.2501-2524
Keywords: Cutting conditions; Artificial neural network; CNC machine; Surface quality; Energy consumption; Processing timeDOI: 10.15376/biores.17.2.2501-2524
Contact information: a: Department of Materials and Material Processing, Artvin Vocational School, Artvin Çoruh University, Artvin, Turkey; b: Department of Forest Industry Engineering, Faculty of Forestry, Karadeniz Technical University, Trabzon, Turkey; c: Department of Wood Mechanics and Technology, Faculty of Forestry, Düzce University, Düzce, Turkey;
* Corresponding author: umitbuyuksari@duzce.edu.tr
INTRODUCTION
The furniture industry is a sector in which solid wood and wood-based panels are consumed in very high quantities to supply a fast-growing market worldwide. In particular, with the use of computer numerical control (CNC) router machines in the furniture industry, production quantities have increased rapidly, while the production costs and labour have decreased (Pelit et al. 2021). These machines are highly preferred in processes such as patterning, milling, drilling, and grooving, and their integration with other automation systems is very flexible. These machines, which increase productivity and reduce time loss, also improve the surface quality of the processed materials (Sutcu 2013; Koc et al. 2017; Sofuoglu 2017).
The most accurate determination of the parameters describing the surface quality of materials such as surface roughness and wettability is extremely significant for the successful application of wood finishing processes such as painting, coating and varnishing (Qin et al. 2015; Sofuoglu 2017; Martha et al. 2020). The most important factors affecting the surface roughness are classified as the factors originating from the material, such as wood species, density, hardness, moisture, and factors originating from the process applied to the material such as cutting tool diameter, tool shape, radius of the tool nose, spindle speed, step-over, depth of cut, and feed rate (Bajić et al. 2008). Wettability, which is sensitive to the interactions between the wood surface and liquid substances such as water, adhesive, paint, varnish, and coating, is traditionally determined by the contact angle, and the small angles indicate more wettable surfaces (Gindl and Tschegg 2002; Gindl et al. 2004; Rathke and Sinn 2013; Fang et al. 2016). The most important factors affecting the wettability of wood surface are wood species, moisture, fibre direction, polarity, pH, surface roughness, wood aging, and processing (Mohammed-Ziegler et al. 2004; Cao et al. 2005; Unsal et al. 2011; Qin et al. 2015).
Table 1. The Comparison of Some Optimization Studies with this Study
Before performing the operations on CNC machines, cutting conditions such as step-over, spindle speed, cutter plunge speed, feed rate, tool diameter, depth of cut, and tool strategy must be set correctly. Otherwise, problems may occur in the surface quality of the processed wood and wood-based panels (Koc et al. 2017; Bal and Akcakaya 2018). The necessity of determining the optimum CNC cutting condition values that give the best surface quality for wood materials used in the furniture industry has recently been an important research issue in the literature. Therefore, many researchers have focused on this issue in their studies. The comparison of some of these studies which investigated on solid wood with this study is presented in Table 1.
Table 1 shows that only surface roughness values are used as output variables in the parameter optimization studies related to CNC-processed solid wood in the literature. Furthermore, it has been determined that factors such as energy consumption and processing time, which are very important in terms of environment and cost, are not used in the optimization studies. Therefore, this study aimed to predict optimum cutting conditions depending on the surface roughness, wettability, energy consumption, and processing time of the solid wood processed with CNC machine.
EXPERIMENTAL
Materials
In this study, walnut (Juglans regia L.) and ash (Fraxinus excelsior L.) wood, which are widely used in the furniture industry, were used unlike the studies in the literature (Table 1). The wood samples were conditioned in an air-conditioning chamber until they reached a moisture content of 12% ± 1% before the cutting process with a CNC machine. A four-axis CNC milling machine (Megatron 2128, Bursa, Turkey) with a spindle power of 9 kW and a maximum spindle speed of 24000 was used for cutting operations. The toolpath strategy was chosen as offset and double flute straight milling cutters in three different diameters (3 mm, 6 mm, and 8 mm) were preferred. The offset strategy was applied constantly in tangential directions of wood samples. The CNC processing of wood samples and the used cutting tools are shown in Fig. 1.
Fig. 1. CNC processing of wood samples and the used cutting tools
Spindle speed and feed rate were determined not only by the most frequently used values in the literature, but also by choosing homogeneous parameter ranges for a successful ANN modelling. Consequently, three spindle speed (12000, 15000, and 18000 rpm) and feed rate (3, 6, and 9 m/min) were used for CNC processing. The depth of cut was 3 mm. The energy consumption of each sample was determined using a wattmeter immediately after processing with the CNC machine, and the total processing times were recorded using a stopwatch. In the energy consumption measurements, a value was taken from the wattmeter just before the CNC process starts, and another value was taken right after the CNC process was finished, and the difference between these two values (last wattmeter value – first wattmeter value) gave the energy consumption of that group. Similarly, the processing time of the groups was determined by using a stopwatch. Afterwards, the samples were sized to 50 mm x 50 mm and test specimens were obtained for surface roughness and wettability measurements. Five test specimens were prepared to represent each group and the test phase was started.
Methods
Surface roughness measurements
The surface roughness measurements were performed according to DIN 4768 (1990) standard to determine the surface quality of wood materials. The Ra (arithmetic mean) values of the measurement parameters were determined perpendicular to the fibres of the wood specimens in the Mitutoyo Surftest SJ-301 test device (Kawasaki, Japan) according to the DIN 4798 (1990) standard. The device with detector nose radius of 5 µm was set as evaluating length of 12.5 mm, cut-off length of 2.5 mm, resolution of 350 µm. Ten measurements were made for each group.
Contact angle measurements
The wettability of the wood surfaces was determined by measuring the contact angles between the surface of the wood specimens and the droplets of distilled water. Using the DSA100 Drop Shape Analysis System (KRUSS GmbH, Hamburg, Germany) equipped with image analysis software, a total of ten drops of 5 µL volume were randomly dropped onto the specimen surfaces. The contact angle values of the specimens were calculated five seconds after the droplets were deposited on the surface.
Artificial neural network analysis
Artificial neural network (ANN) analyses were carried out using the energy consumption and processing time values measured immediately after the CNC process, and the surface roughness and contact angle values obtained from the tests. As a result of the ANN analysis, the prediction models with the best performance were used both to predict the surface roughness, wettability, energy consumption and processing time values of the cutting conditions that were not used in the experimental studies, and to determine the optimum cutting condition values that give the best surface quality and energy-time savings for walnut and ash wood. The CNC cutting conditions such as spindle speed, feed rate, and cutting tool diameter were main variables in ANN modelling of this study. The data obtained from experimental studies were modelled using the MATLAB Neural Network Toolbox. The experimental data were randomly grouped as training data, validation data, and testing data for each test. Training data were presented to the network during training, and the network is adjusted according to its error. Validation data is used to measure network generalization and to halt training when generalization stops improving. Testing data has no effect on training and so provide an independent measure of network performance during and after training. Each network was trained with 38 data (about 70% of total data) and was subsequently validated with 8 experimental data (about 15% of total data) and tested with 8 experimental data (about 15% of total data).
The data sets used in the prediction models are given in Tables 3 and 4. The Levenberg Marquardt algorithm (trainlm) was chosen as the training algorithm. This algorithm typically requires more memory but less time. Training automatically stops when generalization stops improving, as indicated by an increase in the mean square error (MSE) of the validation samples. The MSE calculated by Eq. 1 was preferred as the performance function,
(1)
where ti is the actual output (targeted values), tdi is the neural network output (predicted values), and N is the total number of training patterns.
The feed forward and backpropagation multilayer ANN were used to determine prediction models. In ANN analysis trials, the transfer (activation) function was used the hyperbolic tangent sigmoid function (tansig) in the hidden layer whilst it was used the linear transfer function (purelin) in the output layer. The layers in which these activation functions are used are shown in Fig. 2.
Fig. 2. The network structures of the surface roughness (a), wettability (b), energy consumption (c) and processing time (d) prediction models
Table 2. Connection Weights and Bias Values of the Prediction Models
The actual (measured) values were compared with the prediction values obtained from ANN analyses after the testing process. The performances of the prediction models were determined by using the root mean square error (RMSE) calculated by Eq. 2 and the mean absolute percent error (MAPE) calculated by Eq. 3,
(2)
(3)
where ti is the actual output values, tdi is the neural network predicted values, and N is the number of objects.
The connection weights (w) and biases (b) of the surface roughness and wettability, energy consumption, and processing time prediction models were given in Table 2. Moreover, the neurons and biases were denoted by n and b in Table 2.
RESULTS AND DISCUSSION
Experimental and ANN Analysis Results
The experimentally obtained data and the prediction values obtained from ANN models of these data are given in Tables 3 and 4 according to wood species. In addition, training, validation and testing data sets used in ANN analysis are indicated.
The performance values of the surface roughness, wettability, energy consumption and processing time prediction models are given in Table 5.
The MAPE values for the surface roughness were 4.71% for training, 6.88% for validation, and 9.60% for testing, whilst the values for the wettability were 1.64% for training, 2.02% for validation, and 3.86% for testing. The values for the energy consumption were 1.11% in training phase, 1.62% in validation phase, and 2.33% in testing phase, while the values for the processing time were 1.14% in training phase, 1.97% in validation phase, and 1.53% in testing phase (Table 5). The MAPE value, which is frequently used by researchers to evaluate ANN model performances, is expected to be below 10% (Antanasijević et al. 2013; Tiryaki et al. 2016). It has been demonstrated that the prediction performance of ANN models is high with these values lower than 10% (Yadav and Nath 2017). It is stated in the literature that it is extremely important to calculate RMSE values as well as MAPE values in order to determine the performance of prediction models (Kucukonder et al. 2016). In this study, the RMSE values of the surface roughness prediction model for training, validation, and testing phase were 0.32, 0.42, and 0.45 whilst the values of the wettability were 1.98, 2.23, and 3.98, respectively. The RMSE values of the energy consumption prediction model for training, validation, and testing phase were 1.07, 2.77, and 2.22 while the values of the processing time were 0.37, 1.07, and 0.51, respectively (Table 5). Taspınar and Bozkurt (2014) stated that the low RMSE values obtained from the ANN analyses are an indicator of the successful performance of the prediction models. The MAPE and RMSE values obtained from the study proved that the ANN models used for prediction and optimization are reliable and can give satisfactory accurate results.
The MSE changes of the ANN prediction models depending on iteration are shown in Fig. 3. The best validation performances of the surface roughness, wettability, energy consumption and processing time prediction models were realized in the 6th, 7th, 37th, and 14th iterations, respectively. After these iterations, the training phases of the networks were stopped. The MSE values after this stage are given in Table 5.
Table 3. Experimental and ANN Analysis Results for Walnut Wood
Table 4. Experimental and ANN Analysis Results for Ash Wood