ARMA/USRMS 06-1165

Neural Networks Analysis for Estimating Rock Cuttability from Rock Properties
Bilgin, N.
Prof.Dr., Head of the Mine Mechanization and Technology Division, Istanbul Technical University, Mining Eng. Dept., 34469 Maslak-Istanbul-Turkey, Ph: +90-212-285-6159; Fax: +90-212-285-6131; E-mail: bilgin@itu.edu.tr

Feridunoglu, C., Tumac, D., Copur, H., Balci, C., and Tuncdemir, H.
Istanbul Technical University, Mining Eng. Dept., Mine Mechanization and Technology Division, 34469 MaslakIstanbul-Turkey

ABSTRACT: Twenty-two rock samples having different rock properties are subjected to a set of experimental program in the laboratories of Mining Engineering Department, Istanbul Technical University. In the first stage, rock samples are subjected to a large program of rock mechanics tests. In the second stage, rock samples having size of 40 x 40 x 50 cm are subjected to full-scale rock cutting tests with a type of conical cutter using linear rock cutting machine (LCM) developed under NATO-TU Research program. Some predictor equations using regression analysis are developed to estimate the performance of mechanical excavators utilizing point attack tools. Artificial neural network (ANN) analyses are also performed to see whether it is possible to predict the performance of conical cutters more accurately than statistical analysis. The results indicate that ANN method yields more reliable predictor equations for cutting performance. The ANN models explained in this study will be further refined in future.

1. INTRODUCTION The application of roadheaders in both civil and mining engineering fields has increased significantly in recent years and the prediction of cutter forces has emerged as a necessity to provide basic data for machine selection, design and performance prediction for a given rock formation [1-7]. A number of researchers have studied the theoretical aspects of coal and rock cutting process for the last 40 years. However, the most comprehensive and accepted theories are those of Evans [8-11] for chisel picks and conical picks and of Nishimatsus [12] for chisel picks. Although these theories led to a better understanding of the coal and rock cutting process, the prediction of cutting tool performance have been always felt a necessity, since in some cases theoretical estimations of cutter forces were not in good agreement with actual cutting forces due to the complex petrographical, mineralogical and anisotropic nature of rock formations. Specific energy is one of the most important factors in determining the efficiency of cutting systems and defined as the work to excavate a unit volume of rock. The effect of the spacing between cuts and

depth of cut (or penetration) on cutting efficiency is explained in Fig. 1. If the line spacing is too close (a), the cutting is not efficient because the rock is over-crushed; tool wear is also high in this region due to the high friction between tool and rock. If the line spacing is too wide (c), the cutting is not efficient since the cuts cannot generate relieved cuts (tensile fractures from adjacent cuts cannot reach to form a chip), creating a groove-deepening situation which creates shock loads causing gross failures in cutting edge. The minimum specific energy is obtained with an appropriate spacing to depth of cut ratio (b). One of the most accepted method to predict the cutting rate of any excavation machine is to use cutting power, specific energy obtained in laboratory from full-scale linear cutting test and energy transfer ratio from cutting head to rock formation as in Eq. (1) [2,3]: ICR = k P SE opt (1)

where, ICR is instantaneous cutting rate in m3/h, k is energy transfer ratio, P is cutting power of cutting head in kW and SEopt is optimum specific energy in kWh/m3. It is strongly emphasized that the predicted value of cutting rate is more realistic if specific energy value in Eq. (1) is obtained from

3. ROCK CUTTING TESTS Full-scale cutting rig used in the experiments was designed and manufactured within the NATO-TU Excavation Project with sizeable technical help of the Earth Mechanics Institute of Colorado School of Mines [14]. It can accommodate block rock samples up to 0.7x0.7x1 m in size. A data acquisition system is used to record the cutter forces in three perpendicular directions. The cutter is fixed with a tool holder directly to an aluminum dynamometer The entire test is carried out with an S-35/80H conical pick manufactured by Sandvik. It has a gauge of 80 mm, flange diameter of 64 mm, shank diameter of 35 mm, tip diameter of 22 mm and primary tip angle of 80. The constant conditions throughout the testing program are attack angle of 55, cutting speed of 12.7 cm/s and skew and tilt angles of 0. The data sampling rate is 2000 Hz. The depth of cut is changed usually from 3 to 10 mm and in some cases from 5 to 10 mm. The line spacing is adjusted according to depth of cut to obtain optimum groove interaction. Peak and mean forces (cutting and normal forces) and yield are recorded in each case. Specific energy is obtained by dividing mean cutting force to yield. Yield is defined as the volume of rock obtained per unit cutting length. The relationship between cutter forces and depth of cut is linear within the limits of depth of cut (from 3 to 10 mm) used for all the rocks tested. The linearity between depth of cut and cutter forces are supported by the previous results of Roxborough [15] and Hurt and Laidlaw [16] as seen in Figs. 2 and 3. This linearity enables the defining of a cutting performance parameter of force to depth of cut ratio (kgf/mm) for each experimental rock, which is used in statistical analysis. Cutting test results for relieved and unrelieved cutting modes as explained in Fig. 1 are presented in Table 2. Relieved cutting test results are for optimum cutting conditions at 9 mm depth of cut. The reason for using 9 mm of depth of cut is that the variation of specific energy versus depth of cut in unrelieved cutting tests usually approaches asymptotically to a minimum level at depth of cut values greater than 9 to 10 mm for the tested rocks; after this value of depth of cut (dopt as seen in Fig. 1) the specific energy does not change much. The predictor equations obtained from regression analysis are summarized in Table 3.

Fig. 1. The general effect of cutter spacing on specific energy.

full-scale linear cutting tests at optimum conditions using production cutters. Rostami and Ozdemir pointed out that k changed between 0.45 to 0.55 for roadheaders and from 0.85 to 0.90 for TBMs [2]. An extensive engineering research program on rock excavation is established in the Mining Engineering Department of Istanbul Technical University using 22 different large rock specimens in which the rock cutting mechanics are largely evaluated experimentally. The results are recently published [13]; however, this paper differs from the previous one in a way that artificial neural network is applied for the performance prediction of conical cutters and the predicted values are compared with those obtained by regression analysis techniques. 2. PHYSICAL AND MECHANICAL PROPERTIES OF THE ROCKS TESTED Twenty-two different rock samples are collected from different mining and tunneling sites and subjected to physical and mechanical property tests. The tests are performed on core samples taken perpendicular to bedding planes, if exists. The results are summarized in Table 1.

Table.1. Physical and mechanical properties of the rocks tested Rock Name High-Grade Chromite (*) Medium-Grade Chromite (**) Low-Grade Chromite (***) Copper Ore (Yellow) Copper Ore (Black) Harsburgite Serpantinite Trona Anhydrite Selestite Jips Sandstone-1 Sandstone-2 Sandstone-3 Siltstone Limestone Tuff 1 Tuff 2 Tuff 3 Tuff 4 Tuff 5 Tuff 6 (g/cm ) 4.03 3.39 2.88 4.13 4.07 2.65 2.49 2.13 2.90 3.97 2.32 2.65 2.67 2.67 2.65 2.72 1.49 1.70 1.80 1.71 1.71 1.49
3

UCS Sd (MPa) 32 4.4 47 10.9 46 7.6 33 2.5 41 3.6 58 27.4 38 10.1 30 0.7 82 6.0 29 2.9 33 2.2 114 7.0 174 10.0 87 4.0 58 3.0 121 7.0 10 0.5 11 0.4 27 0.6 14 0.5 19 0.6 6 0.2

BTS Sd (MPa) 3.7 0.6 4.5 0.6 3.7 0.8 3.4 0.02 5.7 0.03 5.5 1.0 5.7 0.5 2.2 0.4 5.5 0.8 4.0 0.5 3.0 0.2 6.6 0.3 11.6 0.4 8.3 0.3 5.3 0.2 7.8 0.3 0.9 0.01 1.2 0.01 2.6 0.02 1.5 0.1 2.3 0.02 0.2 0.01

Esta (GPa) 3.5 2.3 2.9 2.1 2.3 3.4 11.0 17.0 28.0 33.3 30.0 57.0 1.1 1.4 2.4 1.6 1.3 0.4

Edyn (GPa) 31.2 76.4 35.2 42.0 49.6 16.1 13.9 3.7 36.5 62.2 55.0 48.8 37.9 3.8 5.2 7.5 5.2 6.1 2.8

SHRV 30 43 43 52 39 53 57 52 48 55 28 42 31 39 38 14

(*): (46 to 50% Cr2O3), (**): (42 to 46% Cr2O3), (***): (20 to 25% Cr2O3), Sd = Standard Deviation, = Density, UCS = Uniaxial Compressive Strength, BTS = Brazilian Tensile Strength, Esta = Static Elasticity Modulus, Edyn = Dynamic Elasticity Modulus, SHRV = Schmidt Hammer Rebound Value using N-24 Type Hammer.

1000 Mean Cutting and Normal Forces ( kgf ) Coal, Mean Cutting Force Coal, Mean Normal Force Cutter: Conical Pick 600

2100 1800 Mean Cutting Force ( kgf ) 1500 1200 900 600 300 0 0 10 20 30 40 50 0 5 10 15 Depth of Cut ( mm ) 20 25 30 Depth of Cut ( mm ) Grindleford Sandstone (Compressive Strength = 45 MPa) Cutter: Conical Pick

800

400

200

Fig. 2. The relationship between cutter forces and depth of cut during shearing of a continuous miner, (after Roxborough at al. 1981, [15]).

Fig. 3. The relationship between cutter forces and depth of cut, (after Hurt and Laidlaw 1979, [16]).

Table 2. Cutting characteristics of the rocks tested

Rock Name

Unrelieved Cutting Sd (mean at different d values)

Relieved Cutting Sd (at Optimum ( s / d ) Ratio)

FC/d FN/d FC/d FN/d Optimum SEopt kgf/mm kgf/mm kgf/mm kgf/mm (s/d) ratio (kwh/m3) High-Grade Chromite 3 54 2 41 8 40 6 27 6 3.9 0.8 Medium-Grade Chromite 2 81 17 64 4 52 11 38 10 6.4 1.3 Low-Grade Chromite 3 69 7 60 5 51 7 40 5 5.0 1.3 Copper Ore (Yellow) 4 43 11 30 7 40 5 26 4 3.7 0.6 Copper Ore (Black) 4 72 17 77 13 81 7 77 5 9.2 0.9 Harsburgite 5 104 3 115 13 101 10 105 11 8.4 2.0 Serpantinite 3 69 14 78 19 49 6 54 7 6.2 1.3 Trona 3 37 13 57 20 25 5 33 6 2.7 0.6 Anhydrite 5 75 11 85 16 56 6 57 6 3.8 0.5 Selestite 3 43 7 33 3 33 4 25 4 3.0 0.4 Jips 3 36 7 23 9 23 2 15 3 3.4 0.5 Sandstone-1 2 113 16 116 29 89 5 90 7 12.6 1.2 Sandstone-2 2 154 30 199 18 104 8 128 14 15.4 1.1 Sandstone-3 3 66 7 69 7 54 5 46 5 5.4 0.5 Siltstone 3 105 17 132 37 70 6 80 15 9.6 0.7 Limestone 5 129 5 229 23 129 5 192 8 12.0 1.4 Tuff 1 3 4 0.3 1.6 0.02 7 0.6 15 2 10 1.7 Tuff 2 3 16 0.9 2.7 0.03 24 1.4 37 3 25 3 Tuff 3 3 17 0.8 2.2 0.06 24 1.3 29 3 20 0.1 Tuff 4 3 12 0.4 2.4 0.02 18 0.5 22 2 15 0.8 Tuff 5 3 14 0.6 2.1 0.02 23 0.7 31 3 20 1 Tuff 6 3 6 0.2 1.3 0.01 11 0.4 12 2 7 0.8 s = Line Spacing, d = Depth of Cut, FC = Mean Cutting Force, FN = Mean Normal Force, SEopt = Specific Energy obtained at Optimum Cutting Conditions, Sd = Standard Deviation

Table 3. Regression equations obtained from classical statistical method to predict cutter performance Unrelieved Cutting Regression Equation FC/d = 0.826 c + 21.76 FC/d = 12.625 t + 8.78 FC/d = 4.542 e FC/d = 12.278 FC/d = 28.974 FN/d = 15.74
0.058 SH

Relieved Cutting ( at optimum s/d) R

2

Regression Equation FC/d = 2.347 c0.785 FC/d = FC/d = FC/d = FN/d = FC/d = 3.292 e 16.794 t0.721 0.058 SH 8.485 Edyn0.557 21.051 Esta0.432 10.687 t0.947 0.079 SH
0.596

R2 0.808 0.754 0.716 0.646 0.595 0.817 0.735 0.744 0.640 0.527 0.760 0.743 0.757 0.682 0.617

0.810 0.797 0.772 0.652 0.614 0.843 0.760 0.784 0.683 0.524

Edyn0.519 Esta0.413

FN/d = 1.217 c1.014 FN/d = 1.723 e FN/d = 8.236 t0.915 0.079 SH

0.589

FN/d = 0.752 c1.051 FN/d = 1.141 e FN/d = 5.266

FN/d = 20.893 Esta

FN/d = 14.136 Esta

Edyn0.633

Edyn0.668

FC = Mean Cutting Force in kgf, FN = Mean Normal Force in kgf, d = Depth of Cut in mm, SEopt = Specific Energy at Optimum Cutting Condition in kWh/m3, c = Uniaxial Compressive Strength in MPa, t = Brazilian Tensile Strength in MPa, Esta = Static Elasticity Modulus in GPa, Edyn = Dynamic Elasticity Modulus in GPa, SH = N-24 Type Schmidt Hammer Rebound Value.

SEopt = 0.083 c + 1.424 SEopt = 1.259 t + 0.142 SEopt = 0.3912 e SEopt = 0.984
0.058 SH

Edyn0.542

SEopt = 2.424 Esta0.414

Regression analysis results in that the best correlations are obtained for uniaxial compressive and Brazilian tensile strength values suggesting that these are the most important rock properties affecting the performance of conical picks. The third dominant rock property is found to be Schmidt hammer rebound value obtained from N-24 type hammer. Static and dynamic elasticity modulus values yield lower correlations. The predictor equations given in Table 3 may enable any engineer to calculate tool forces from rock properties within acceptable statistical limits. The estimation of optimum specific energy is important in predicting cutting rates of excavation machines as explained before. As seen from Table 3, optimum specific energy is best predicted from uniaxial compressive strength and tensile strength verifying some of the previously published results [17, 18]. Moderate correlations are obtained for Schmidt hammer rebound values for N-24 type hammer and static and dynamic elasticity modulus. Cutter spacing to depth of cut ratio (s/d) is the key factor in obtaining optimum specific energy, hence the most efficient cutting conditions. A direct relationship between rock properties and optimum s/d ratio could not be found within the limits of this research program. It is strictly recommended that this aspect of rock cutting mechanics needs further research. 4. ARTIFICIAL NEURAL NETWORKS (ANN) Artificial neural networks (ANN) are simplified models of the biological structure in human brains, its development started as an attempt to understand the operation of human brain and to decide and act under uncertainty. The basic terminology is borrowed from neuroscience. The models based on ANN consisting of elementary processing units called neurons. Large amount of interconnection between neurons and their capability to learn from data set serve to solve complex problems as a strong tool for prediction. In ANN, the mathematical relationships between different variables are not specified, instead they learn from examples. Back propagation learning algorithm is used in this study; this consists of a number of interconnected processing units, referred as artificial neurons. The neurons are arranged into two or more layers and interact with each other via weighed connections. The weights determine the nature and strength of

the influence between the interconnected neurons. Each neuron is connected to all the other neurons in the next layer. The basic units of neural networks, the artificial neurons, simulate the four basic functions of natural neurons. Fig. 4 shows a fundamental representation of an artificial neuron.

Fig. 4. A basic artificial neuron.

In Fig. 4, various inputs to the network are represented by the mathematical symbol, xn. Each of these inputs is multiplied by a connection weight. These weights are represented by wn. In the simplest case, these products are simply summed, fed through a transfer function to generate a result, and then output. This electronic implementation is still possible with other network structures, which utilize different summing functions, as well as different transfer functions. Artificial neural networks, with their remarkable ability to derive meaning from complicated or imprecise data, can be used to detect trends form complex data sets. A trained neural network can be thought of as an "expert" in the category of information it has been given to analyze. Typical neural network structure is shown in Fig. 5. Layers have neurons (i.e. input processing point) and every neuron have a connection with other neurons. Neuron connections are identified as weight (w) parameters. Input is processed with weights through neurons with activation function and forms output. Neuron output is calculated by activating the previous layer neurons with an activation function. The most widely used activation function is sigmoid and it is given in Eq. (2):

x0 =

1 1 + exp( xh who )

(2)

where xh is the activation of the hth neuron in previous layer and who is the weight between two neurons.
W12 X1 W13 W25 W35 Y1

Wij

Wjk

Input Layer

Hidden Layer

Output Layer

Fig. 5. Typical neural network structure used in this study.

In order to train a neural network to perform some task, the weights of each unit must be adjusted in such a way that the error between the desired output and the actual output is reduced. In other words, the error changes must be calculated as each weight is increased or decreased slightly. The back propagation (BP) algorithm is the most widely used method for training neural network systems. At the beginning of the training, initial weights are given randomly between 0.1 and 0.1. As the training starts, the weights are calculated up to a point where the error between input and output is reached at a certain desired value. Artificial neural networks have been widely used for automatic control, signal processing and optimization. Gajewsky and Jonak in their recent studies showed that wear of cutting tools in a mechanical excavator can be controlled with the ANN using a signal processing system and a more efficient cutting process may be obtained [19, 20]. ANN has found also a wide application in rock mechanics and geotechnical area. Yang and Zhang showed that neural networks can learn about rock engineering systems, to incorporate rock mechanics parameters and even be able to link with environmental and financial aspects [21]. Meulenkamp and Grima applied the neural networks for the prediction of the uniaxial compressive strength from equotip hardness [22]. Singh and co-authors conclude that prediction of

rock strength properties from petrographical properties using ANN was more accurate than conventional statistical models [23]. Benardos and Kaliampakos demonstrated that ANN could be easily used for TBM performance estimation and projects strategic developments [24]. Sonmez and his colleagues emphasized that sometimes it was difficult to obtain standard core samples from weak, stratified, highly fractured rock masses. In such cases, ANN could be used with a great precision as a prediction tool [25]. Javadi in his recent study showed the capabilities of ANN to predict air loss from tunnels driven under compressed air. A feed forward back propagation neural network was trained and used to predict the air loss in compressed air tunneling. The results of the neural network training and predictions were compared with the field measurements. It was also shown that, a neural network could learn from not only the relationship between appropriate soil and tunnel parameters and air losses, it could also generalize the learning to predict air loss for different geological conditions [26]. Kahraman and his friend proved that the prediction of slap production in marble industry could be more accurately predicted using ANN models than using statistical models [27]. 5. ARTIFICIAL NEURAL NETWORKS USED IN THIS STUDY Multi layered perception models consisting of 3 layers (input x hidden x output) are used in this study to estimate cutting force, normal force and optimum specific energy obtained from rock cutting experiments, based on mechanical properties such as uniaxial compressive strength, Brazilian tensile strength, Schmidt hammer rebound value, dynamic elastic modulus and static elastic modulus. The neural network models used in this study are implemented in MATLAB. All models have one neuron in the input and in the output layers. The number of neurons in the hidden layer varies with complexity of the data set and trials are performed to find suitable number of hidden layer neurons, which reacts best in the training process. Neural network structures as numbers of neurons within the layers (input x hidden x output) are presented in Table 4. Back propagation training algorithm is used in all models. Training epoch is set to 500 for the training process.

Table 4. Artificial neural network structures applied to data Input Parameter UCS BTS SHRV Edyn Esta UCS BTS SHRV Edyn Esta Output Parameters Unrelieved FC 1 x 10 x 1 1 x 15 x 1 1 x 8 x1 1 x 50 x 1 1 x 50 x 1 Unrelieved FN 1 x 10 x 1 1 x 15 x 1 1 x 8 x1 1 x 50 x 1 1 x 50 x 1 Relieved FC 1 x 10 x 1 1 x 10 x 1 1x8x1 1x8x1 1 x 30 x 1 Relieved FN 1 x 25 x 1 1 x 10 x 1 1x8x1 1 x 40 x 1 1 x 50 x 1 SEopt 1 x 50 x1 1x2x1 1 x 50 x1 1 x 50 x1 1 x 50 x1

Table 5. Root mean square error values for artificial neural network and statistical models Predictor Parameter UCS BTS SHRV Edyn Esta RMSE Values for Predicted Parameters Unrelieved FC Relieved FC ANN Statistical ANN Statistical Model Model Model Model 11.06 16.28 16.07 17.60 14.72 16.83 15.25 18.60 6.19 19.86 4.49 17.71 16.67 27.77 11.86 26.24 4.29 27.84 21.28 23.11 Unrelieved FN Relieved FN ANN Statistical ANN Statistical Model Model Model Model 15.69 27.76 23.59 26.39 14.06 36.15 25.07 30.77 9.10 34.11 24.58 29.79 14.78 90.59 29.86 41.94 24.16 59.62 15.80 29.71 SEopt ANN Model UCS BTS SHRV Edyn Esta 0.28 1.53 0.31 0.19 0.39 Statistical Model 1.90 1.97 2.24 2.98 2.77

UCS = Uniaxial Compressive Strength, BTS = Brazilian Tensile Strength, SHRV = Schmidt Hammer Rebound Value using N24 Type Hammer, Edyn = Dynamic Elasticity Modulus, Esta = Static Elasticity Modulus, FC = Mean Cutting Force, FN = Mean Normal Force, SEopt = Optimum Specific Energy

UCS BTS SHRV Edyn Esta

Root mean square error (RMSE) is utilized for evaluation of model performance. The smaller the RMSE, the better the performance of the model. RMSE values are calculated as in Eq. (3):
1 n RMSE = y y i i n i =0 2

(3)

The predicted performance parameters of conical cutter obtained from ANN models with the input parameter of uniaxial compressive strength are plotted against measured values as shown in Fig. 610. As seen from these figures and consulting the RMSE values given in Table 5, it may be concluded that trained data using ANN model gives much reliable predicted values compared to classical statistical analysis methods.

ANN Output Unrelieved Cutting Force / Depth of Cut (kgf/mm)

where n is number of data, yi is observed data and i is the model output. RMSE values for ANN models and statistical models are presented in Table 5. As seen from this table, for each predictor parameter ANN model gives the smaller RMSE compared to statistical model.

RMSE = Root Mean Square Error, UCS = Uniaxial Compressive Strength, BTS = Brazilian Tensile Strength, SHRV = Schmidt Hammer Rebound Value using N24 Type Hammer, Edyn = Dynamic Elasticity Modulus, Esta = Static Elasticity Modulus, FC = Mean Cutting Force, FN = Mean Normal Force, SEopt = Optimum Specific Energy

Input: Uniaxial Compressive Strength Output: Unrelieved Cutting Force 200

150

100

50

y = 0,913x + 5,5211 R2 = 0,913

0 0 50 100 150 200 Measured Unrelieved Cutting Force / Depth of Cut (kgf/mm)

Fig. 6. Measured unrelieved cutting force versus ANN model output with input parameter of uniaxial compressive strength.

Input: Uniaxial Compressive Strength Output: Relieved Cutting Force 200 ANN Output Optimum Specific Energy (kWh/m 3) ANN Output Relieved Cutting Force / Depth of Cut (kgf/mm) 20

Input: Uniaxial Compressive Strength Output: Optimum Specific Energy

150

15

100

10

50

y = 0,7531x + 12,389 R2 = 0,7531

5 y = 0,9949x + 0,0285 R2 = 0,9949 0 5 10 15 20

0 0 50 100 150 200 Measured Relieved Cutting Force / Depth of Cut (kgf/mm)

Measured Optimum Specific Energy (kWh/m3)

Fig. 7. Measured relieved cutting force versus ANN model output with input parameter of uniaxial compressive strength.

Fig. 10. Measured optimum specific energy versus ANN model output with input parameter of uniaxial compressive strength.

Input: Uniaxial Compressive Strength Output: Unrelieved Normal Force ANN Output Unrelieved Normal Force / Depth of Cut (kgf/mm) 200

6. CONCLUSIONS The results of experimental studies indicate that uniaxial compressive strength and tensile strength are best correlated with the measured cutting and normal force values and optimum specific energy values in statistical regression analyses. These equations from regression analyses may enable engineers to calculate tool forces for relieved and unrelieved conditions from rock properties within acceptable statistical limits. Multi layered perception models consisting of 3 layers and back propagation training algorithm is used also to predict conical cutter performance. It may be concluded that trained data using ANN model gives much reliable predictions compared to classical statistical analysis methods. ACKNOWLEDGEMENTS This paper is a summary of a research program carried out in Istanbul Technical University, Mining Engineering Department during the last 8 years. The research was sponsored by NATO Sfs program, the Turkey Republic Prime Ministry State Planning Organization (DPT), Istanbul Technical University Research Fund, Karaelmas University Research Fund and General Directorate of Mineral Research and Exploration (MTA). The authors acknowledge Prof.Dr. Levent Ozdemir from the Earth Mechanics Institute of Colorado School of Mines and all the research staff involved without their help and work, this contribution to rock cutting mechanics could not be possible.

150

100

50 y = 0,9269x + 4,9977 R2 = 0,9269

0 50 100 150 200 Measured Unrelieved Normal Force / Depth of Cut (kgf/mm)

Fig. 8. Measured unrelieved normal force versus ANN model output with input parameter of uniaxial compressive strength.

Input: Uniaxial Compressive Strength Output: Relieved Normal Force 200 ANN Output Relieved Normal Force / Depth of Cut (kgf/mm)

150

100

50 y = 0,7286x + 13,597 R2 = 0,7286

0 50 100 150 200 Measured Relieved Normal Force / Depth of Cut (kgf/mm)

Fig. 9. Measured relieved normal force versus ANN model output with input parameter of uniaxial compressive strength.

REFERENCES
[1] Rostami, J. and L. Ozdemir. 1994. Roadheader performance optimization for mining and civil construction. In Proceedings of the 13th Annual Technical Conference, Institute of Shaft Drilling Technology, Las Vegas, Nevada. eds. DeMers JE, et al. [2] Rostami, J., L. Ozdemir, and D.M. Neil. 1994. Performance prediction: a key issue in mechanical hard rock mining. Mining Engineering. Nov.:1264-1267. [3] Neil, D.M., J. Rostami, L. Ozdemir, and R. Gertsch. 1994. Production estimating techniques for underground mining using roadheaders. In Society for Mining, Metallurgy and Exploration (SME) Annual Meeting, Preprint No: 94-233. pp. 1-7. [4] Copur, H., J. Rostami, L. Ozdemir, and N. Bilgin. 1997. Studies on performance prediction of roadheaders based on field data in mining and tunnelling projects. In Proceedings of the 4th International Symposium on Mine Mechanization and Automation, Brisbane, Queensland. eds. Gurgenci, H. and M. Hood. p. A4-1/A4-7. [5] Cigla, M., S. Yagiz, and L. Ozdemir. 2001. Application of tunnel boring machines in underground mine development. In Proceedings of the 17th International Mining Congress and Exhibition of Turkey, Ankara. Eds. Unal, E., E. Unver, and E. Tercan. pp. 155-164. [6] Asbury, B., L. Ozdemir, and T. Rozgonyi. 2001. Frustum bit technology for continuous miner and roadheader applications. In Proceedings of the 6th International Symposium on Mine Mechanization and Technology, Johannesburg. ed. Willise R.P.H. p. 135-139. [7] Asbury, B., M. Dezeeuw, M. Cigla, and L. Ozdemir. 2003. Results of practical design modifications for respirable dust reduction on continuous miners in underground coal mining. In Society for Mining, Metallurgy and Exploration. [8] Evans, I. and C.D. Pomeroy. 1966. The strength, fracture and workability of coal. Pergamon Press, Library of Congress Catalog Card Nr. 66-14657. p. 277. [9] Evans, I. 1972. Line spacing of picks for efficient cutting. Int J Rock Mech Min Sci. 9: 355-359. [10] Evans, I. 1984. A theory of the cutting force for point attack picks. International Journal of Mining Engineering. 2: 63-71. [11] Evans, I.1984. Basic mechanics of the point attack pick. Colliery Guardian. May: 189-193. [12] Nishimatsu, Y. 1972. The mechanics of the rock cutting. Int J Rock Mech Min Sci. 9: 261-271. [13] Bilgin, N., M.A. Demircin, H. Copur, C. Balci, H. Tuncdemir, and N. Akcin. 2006. Dominant rock properties affecting the performance of conical cutters and the comparison of some experimental and theoretical results. Int J Rock Mech Min Sci. 43: 139156. [14] Eskikaya, S., N. Bilgin, L. Ozdemir, et al. 2000. Development of rapid excavation technologies for the Turkish mining and tunneling industries. NATO TUExcavation SfS Programme Project Report. Istanbul Technical University, Mining Eng. Dept., Preparation of Report by N. Bilgin and C. Balci, Sept, 2000. p. 172. [15] Roxborough, F.F., P. King, E.J. Pedroncelli . 1981. Tests on the cutting performance of a continuous miner. J S Afr Inst Min Metall. 81: 926.

[16] Hurt, K.G. and D.G. Laidlaw. 1979. Laboratory comparison of three rock-cutting tools. Tunnels and Tunnelling. 11: 13-16. [17] Copur, H., H. Tuncdemir, N. Bilgin, and T. Dincer. 2001. Specific energy as a criterion for the use of rapid excavation systems Turkish mines. Trans Inst Min Metall, Section A. 110: A149-A157. [18] Copur, H., N. Bilgin, H. Tuncdemir, and C. Balci. 2003. A set of indices based on indentation tests for assessment of rock cutting performance and rock properties. J S Afr Inst Min Metall. 103 (9): 589-600. [19] Gajewski, J. and J. Jonak. 2005. Utilization of neural networks to identify the status of the cutting tool point. Tunneling and Underground Space Technology. 21: 180184. [20] Jonak, J. and J. Gajewski. 2006. Identify the cutting tool type used in excavations using neural networks. Tunneling and Underground Space Technology. 21: 185189. [21] Yang, Y. and Q. Zang. 1998. The application of neural networks to rock engineering (RES). Int J Rock Mech Min Sci. 35: 727-745. [22] Meulenkamp, F. and M.A. Grima. 1999. Application of neural networks for the prediction of the unconfined compressive strength (UCS) from equodip hardness. Int J Rock Mech Min Sci. 36: 29-39. [23] Singh, V.K., D. Singh, and T.N. Singh. 2001. Prediction of strength properties of some schistose rocks from petrographic properties using artificial networks. Int J Rock Mech Min Sci. 38: 269-284. [24]Bernardos, A.G. and D.C. Kaliampakos. 2004. Modeling TBM performance with artificial neural networks. Tunneling and Underground Space Technology. 19: 597604. [25] Sonmez, H., C. Gokceoglu, H.A. Nefeslioglu, and A. Kayabas. 2006. Estimation of rock modulus for intact rock with an artificial neural network and for rock masses with a new empirical equation. Int J Rock Mech Min Sci. 43: 224-235. [26] Javadi, A.A. 2006. Estimation of losses in compressed air tunneling using neural network. Tunneling and Underground Space Technology. 21: 9-20. [27] Kahraman, S., H. Altun, B.S. Terekici, and M. Fener. 2006. Sawability prediction of carbonate rocks from shear strength parameters using artificial neural networks. Int J Rock Mech Min Sci. 43: 157-164.