(Microsoft Word - 67-76\343\345\344\317) 1 Al-Khwarizmi Engineering Journal Al-Khwarizmi Engineering Journal, Vol. 14, No. 1, March, (2018) P.P. 67-76 Neural Network Modeling of Cutting Force and Chip Thickness Ratio For Turning Aluminum Alloy 7075-T6 Mohanned Mohammed H. AL-Khafaji Department of Production Engineering and Metallurgy/ University of Technology/ Baghdad/ Iraq Email: mohannedalkhafaji@hotmail.com (Received 15 March 2017; accepted 31 October 2017) https://doi.org/10.22153/kej.2018.10.004 Abstract The turning process has various factors, which affecting machinability and should be investigated. These are surface roughness, tool life, power consumption, cutting temperature, machining force components, tool wear, and chip thickness ratio. These factors made the process nonlinear and complicated. This work aims to build neural network models to correlate the cutting parameters, namely cutting speed, depth of cut and feed rate, to the machining force and chip thickness ratio. The turning process was performed on high strength aluminum alloy 7075-T6. Three radial basis neural networks are constructed for cutting force, passive force, and feed force. In addition, a radial basis network is constructed to model the chip thickness ratio. The inputs to all networks are cutting speed, depth of cut, and feed rate. All networks performances (outputs) for all machining force components (cutting force, passive force and feed force) showed perfect match with the experimental data and the calculated correlation coefficients were equal to one. The built network for the chip thickness ratio is giving correlation coefficient equal one too, when its output compared with the experimental results. These networks (models) are used to optimize the cutting parameters that produce the lowest machining force and chip thickness ratio. The models showed that the optimum machining force was (240.46 N) which can be produced when the cutting speed (683 m/min), depth of cut (3.18 mm) and feed rate (0.27 mm/rev). The proposed network for the chip thickness ratio showed that the minimum chip thickness is (1.21), which is at cutting speed (683 m/min), depth of cut (3.18 mm) and feed rate (0.17 mm/rev). Keywords: Machining forces, chip thickness ratio, neural network, optimization, turning operation. 1. Introduction The turning process is among the most significant cutting operation. It would once generate a variety of cylindrical products like solid, hollow, profile shafts and threads, etc. Due to its important, a lot of scientists considered the parameter which impacting the process either to generate a good finished product, improve tool life or both. Additionally, they examined the power usage reduction and the production time [1]. The machining force ���� in turning operation is a three-dimensional vector. Three components represent it, namely, the cutting force��� ) which is in the direction of cutting axis, the passive force ���� in the direction of radial axis and feed force ��� � in the direction of feed axis as shown in Fig. 1. The cutting force has the biggest value in the three force components. Several researchers learned such components and taking into accounts the effect of cutting variables. Stachurski, et al. [2] utilized a power polynomial to model the cutting force during turning steel C45. Astakhov and Xiao [1] applied mathematical models to estimate the cutting forces during machining two materials, aerospace aluminum alloy 2024 and T6AISI bearing steel E52100. Agustina, et al. [3] implemented a design of experiment to evaluate the impact of cutting factors to the cutting force when turning aluminum alloy (UNS A97075) in dry conditions. They examined the influence of micro grove size and shape on the cutting temperature, cutting force and tool wear. C.X.Yue, et al. [4] produced a three-dimensional model by using Mohanned Mohammed H. Al-Khwarizmi Engineering Journal, Vol. 14, No. 1, P.P. 67- 76 (2018) 68 Abaqus/Explicit to simulate the cutting operation of hardened steel GCr15. For their model the cutting temperature, surface residual stresses, cutting force and the side flow were investigated. The chip thickness ratio (CTR) gives essential indication for the cutting process stability. It can be explained as the ratio relating the chip thickness to the undeformed chip thickness. It is usually greater than unity (CTR>1) [5]. Through the definition, the higher CTR means that the chip is thicker. The reason is the limitation to the chip movement, that in turns, can cause rise in the machining power and vice versa. Santos, et al. [5] researched the machining force (Fu), chip thickness ratio (CTR) and chip disposal during turning ductile (1350-O grade) and high strength (7075-T6 grade) aluminum alloys at different cutting conditions. Astakhov and Shvets [6] investigated the chip compression ratio with several cutting parameters. Fig. 1. Machining force and its components. In recent years, the scientific approaches such as neural network, fuzzy logic, genetic algorithm, ant colony or combinations of them, are used to model nonlinear, complicated and multi parameters system. In addition, they are used in the optimization of such systems. The neural network is miming human brain. It consists of an input layer to presents data to the network, output layer to produces the network response, and one or more hidden layers. The hidden and output layers’ topology, weights and activation functions are the network characterization. A neural network is trained with various data sets and tested with other testing data sets to reach an optimum topology and weights. Once the network is trained, it can be used for prediction, simulation, monitor and control complicated system [7]. Sick [8], used the neural network to estimate the development of tool wear. Sharma, et al., [9] utilize the neural network to model the cutting force and surface roughness as a response to the approach angle, cutting speed, feed rate and depth of cut. Chen, et al., [10] constructed nested artificial neural network. Their model consists from two networks, the first one is the enclosed network which take the cutting parameters to predict the cutting force and tool vibration, and the second is the output network which take the outputs of the first network and the cutting parameters as inputs and give the surface roughness as output. Sangwana, et al., [11] optimized the surface roughness during turning of Ti-6Al-4V titanium alloy by integrating feed forward neural network and the genetic algorithm. AL-Khafaji, et al., [12] applied Levenberg- Marquardt algorithm for backpropagation training algorithm to train four feed forward neural network. Their networks were constructed to different insert type. The network takes the cutting speed, feed rate and depth of cut as input and predict the surface roughness. These networks are used to optimize the cutting parameters for minimizing surface roughness. Mia and Dhar, [13] presented an artificial neural network based model to predict the surface roughness of EN 24T steel in turning operation. Their model take the cutting speed, feed rate, material hardness and the machining environment, coolant or dray conditions, as input. The model output was surface roughness. This paper aims to build neural network model to correlate the cutting variables, cutting speed (�� ), depth of cut ( ), and feed rate ( ), to the machining force (��) and the chip thickness ratio during machining aluminum alloy 7075-T6. 2. Experimental Data The implemented experimental data are conducted by Santos, et al. [5]. The workpieces are artificially aged aluminum alloy 7075-T6, they are cylindrical extruded bars (Ø 101×2,000 mm) in dimension. Their chemical composition is 1.20– 2.00 % Cu, 0.40 % Si, 2.10–2.90 % Mg, and 5.10– 6.10 % Zn. The experiments had been executed on CNC lath machine ROMI Multiplic 35D applying 6% concentration of soluble oil with 360 l/h. The cutting tool implemented comes with ISO designation of (TCGT16T308-AZ HTi10) which is cemented carbide inserts. The tool holder utilized in the experiments is made by Mitsubishi which has a designation of (STGCR2020K16Z). The applied tool geometries have been: rake angle, �� 15°; Mohanned Mohammed H. Al-Khwarizmi Engineering Journal, Vol. 14, No. 1, P.P. 67- 76 (2018) 69 relief angle, �� 7°and approach angle, �� 90°. These angles have been estimated after installing the tool on the tool holder. The forces measuring system is made up from three elements, a force sensor which is force dynamometer, a signal conditioning and USB 6251 data acquisition board. The force sensor and a signal conditioning element are made by Kistler company both have model no. (9265B) and (5019B), respectively. The USB 6251 data acquisition board made by National Instruments controlled by LabVIEW® 9.0 software were applied for data recording. When the cutting conditions are getting a steady-state stage, the data recorded for a 10s interval at 6kHz as sample rate. The system has been calibrated prior to conducting the experiments. The machining variables that will be considered in this paper are cutting speed (�� � depth of cut � � and feed rate � ). Five level were given for each variable, for cutting speed, �� , (117, 200, 400, 600, and 683 m/min), for depth of cut, , (0.38, 1.00, 2.50, 4.00, and 4.62 mm) and for feed rate, , (0.170, 0.200, 0.275; 0.350; and 0.380 mm/rev). The experimentation output were cutting force, �� , passive force, ��, feed force, �� , and chip thickness ratio, ���.The experimental data shown in table 1. The tests no. 8, 9, 10 and 11 shown in the table1 duplicated so that average of their results has been utilized in the modeling. 3. Neural Netowrk Modelling The feedforward networks have number of neurons in their layers, the layers arrange sequentially. The outputs of one layer are inputs to the next layer neurons. As mentioned in advance that the feedforward neural network consists from one or more hidden layer. These layers are characterized by their activation function and neurons number [13]. The network training is a process to adjusts the networks’ weights to reach the minimum error between the network output and the target, the experimental data. The most common algorithm used to train neural network, adjusting weights, is the Backpropagation algorithm. [7]. Table 1, Machining experimental results taken from Santos, et al. [5] No Input Measured �� � ��� � ! (mm) " � �� #$% � &� �'� &(�'� &"�'� )*+ 1 117 2.5 0.275 564 -24.9 158 1.32 2 200 4 0.2 749 -1.9 233 1.69 3 200 4 0.35 1150 -23.1 222 1.71 4 200 1 0.2 167 -5.76 31.5 1.5 5 200 1 0.35 257 -11.4 30.5 1.96 6 400 4.6 0.275 923 -40.4 134 1.45 7 400 2.5 0.17 377 -9.5 133 1.48 8 400 2.5 0.275 518 -30.7 112 1.45 9 400 2.5 0.275 518 -31 113 1.45 10 400 2.5 0.275 522 -32.7 114 1.45 11 400 2.5 0.275 520 -27.9 115 1.45 12 400 2.5 0.38 636 -51.1 101 1.41 13 400 0.38 0.275 84.8 28.6 14 1.14 14 600 4 0.2 678 4.51 185 1.5 15 600 4 0.35 992 -28 167 0.75 16 600 1 0.2 153 -8.64 24.9 1.5 17 600 1 0.35 238 -19.1 11.8 1.61 18 683 2.5 0.275 491 -29.5 106 1.36 The radial basis function (RBF) neural networks type is fundamental categories of neural networks. The primary features of the RBF model are its efficiency, the implementation simplicity. In addition, good learning and generalization capabilities. The radial basis function network construction requires two different layers, a single hidden layer and the output layer. The hidden layer Mohanned Mohammed H. Al-Khwarizmi Engineering Journal, Vol. 14, No. 1, P.P. 67- 76 (2018) 70 has nonlinear processing neuron, which provides an alternative goal from that in the feedforward multilayer perceptron MLP network. The output layer has neurons to compute the scalar product of its inputs and provides the response of the network. The input space transformation to the hidden-unit space is nonlinear, whereas it is linear from the hidden-unit space to the output space. It can be concluded that the RBF network is a feedforward neural network with single hidden-layer [14]. The RBFs are generally proven to have universal approximation capabilities. They are suitable for solving pattern classification and function approximation problems because of their uncomplicated topology and their capability to show the learning proceeds in an explicit manner [14]. The hidden layer activation function in the radial basis neural network is radial function. The most radial basis function used is Gaussian function. In a RBF network having k radial units in the intermediate layer and one output [15]. The weights connecting the hidden and output units are estimated either by the least mean square (LMS) or the gradient descent method [14]. Radial basis networks might need more neurons compared to standard feedforward backpropagation networks, although they can be designed with a less time that it takes to train standard feedforward networks. They operate most effective when many training vectors are implemented [16]. In this work a RBF neural network were used to model the cutting parameters against machining force components and chip thickness ratio. Four models were constructed using MATLAB neural network toolbox. The input to all networks are cutting speed �� , depth of cut and feed rate . The first three networks’ responses are cutting force �� , passive force �� and feed force �� , respectively. Whereas, the fourth network’s response is chip thickness ratio ���. As stated in advance the RBF network is like feedforward MLP network in architecture with only one hidden layer. The function, newrb, in Matlab neural toolbox used in this work to generate the models networks is conducting it calculates the distance of network input from the weights’ matrix rows, rather than matrix multiplication as in MLP network. In addition, it multiplies the bias instead of adding it. Therefore, the input of hidden layer ,-. neuron is computed by (1) [16] [17]. /01 23 − 5012. 701 ... (1) Where, 3 is the input vector, 51 is the weights’ matrix and 71 is the bias. Each weights’ matrix element is considered as center point, a point at which the net input is zero. Whereas, the bias is used to scale the output of the hidden layer transfer function (the radial basis function) output either stretching or compressing it. In this paper the tool box function newrb is implemented to generate the radial basis neural network models. The network generated by this function use the Gaussian function shown in Fig. 2 as a transfer function and defined by (2) 801 9:;<= ...(2) An essential property of the Gaussian function, it is local. Which indicate that the output is near to zero if n moves extremely far in either direction from the center point. In addition, it is global function. It is opposed to the global sigmoid functions used in the multilayer perceptron MLP whose output remain near to 1 when the net input goes toward infinity. The output layer in RBF network is pure line given by (3) [16] [17] 8> ∑ 5@> 8@1A@B1 + 7> …(3) Where, D is the number of neurons in the hidden layer, 5> is the weights’ vector connecting the hidden layer and the output layer and 7> is the bias of the network output layer. The model of the proposed radial basis models is shown in fig. ٣ . The vector 3 in equation (1) is consist of three components which are cutting speed, depth of cut and feed rate. The number of hidden layer neurons (D� is 18 for all four models. The weights matrix 51 has size of 18 rows and 3 columns and the bias vector 71 has 18 elements. The 5> has 18 elements too. Table 2 to table 5 show the network weights matrices and bias for �� , ��, �� and ��� models, respectively. It should be noted that the value of second layer bias 7> for �� , ��, �� and ��� networks are scalar values equal to (−1058.152), (−15.3309), (557.1898) and (0.8283), respectively. Fig. 2. Gaussian function. Mohanned Mohammed H. Al-Khwarizmi Engineering Journal, Vol. 14, No. 1, P.P. 67- 76 (2018) 71 Fig. 3. The proposed neural network model. 4. Results and Discussion The networks outputs are extremely matching the experimental data. The correlation coefficient � is computed for all networks using equation (4) � A ∑ HI ;I:∑ H ∑ ; JKA ∑ HI=:�∑ HI�=LKA ∑ ;I=:�∑ ;I�=L … (4) All networks responses gave value of R equal to one. This is a perfect match. Fig. 4-7) which showing the networks responses compared with the experimental data taken from Santos, et al. [5] for �� , ��, �� , and ���, respectively. It can be seen from those figures that all networks outputs are perfectly coincide with the experimental data given by Santos, et al. [5]. The machining force �� is the resultant of the three components as mentioned in advanced. It can be computed using equation (5). According to the perfect matching between experimental and networks outputs of the machining forces’ components, the machining force �� computed from the networks outputs is perfectly coincide with the experimental ��. Fig. 8 shows the perfect match of the computed �� versus the experimental �� from Santos, et al. [5]. �� J��> + ��> + ��> … (5) Table 2, The &� network weights matrices and bias. no. MN ON MP 1. 200 4.00 0.350 0.833 13961.3 2. 400 2.50 0.380 0.833 0 3. 600 4.00 0.350 0.833 11092.91 4. 400 4.60 0.275 0.833 1913.422 5. 117 2.50 0.275 0.833 1622.152 6. 683 2.50 0.275 0.833 1549.152 7. 200 1.00 0.350 0.833 3520.709 8. 600 1.00 0.350 0.833 3356.444 9. 200 4.00 0.200 0.833 -11940.5 10. 600 4.00 0.200 0.833 -9187.46 11. 400 2.50 0.170 0.833 -8603.13 12. 400 0.38 0.275 0.833 1079.067 13. 200 1.00 0.200 0.833 -2244.6 14. 600 1.00 0.200 0.833 -2096.74 15. 400 2.50 0.275 0.833 9977.41 16. 400 2.50 0.275 0.833 0 17. 400 2.50 0.275 0.833 0 18. 400 2.50 0.275 0.833 0 Table 3, The &( network weights matrices and bias. no. MN ON MP 1. 400 2.5 0.38 0.8326 0 2. 400 4.6 0.275 0.8326 -24.312 3. 400 0.38 0.275 0.8326 44.644 4. 683 2.5 0.275 0.8326 -14.169 5. 400 2.5 0.17 0.8326 1381.82 6. 117 2.5 0.275 0.8326 -9.569 7. 600 1 0.35 0.8326 -335.18 8. 200 4 0.35 0.8326 -683.21 9. 600 4 0.35 0.8326 -1047.9 10. 600 4 0.2 0.8326 1051.56 11. 200 4 0.2 0.8326 686.05 12. 200 1 0.35 0.8326 -177.49 13. 600 1 0.2 0.8326 336.650 14. 200 1 0.2 0.8326 184.293 15. 400 2.5 0.275 0.8326 -1387.3 16. 400 2.5 0.275 0.8326 0 17. 400 2.5 0.275 0.8326 0 18. 400 2.5 0.275 0.8326 0 Mohanned Mohammed H. Al-Khwarizmi Engineering Journal, Vol. 14, No. 1, P.P. 67- 76 (2018) 72 Table 4, The &" network weights matrices and bias no. MN ON MP 1. 200 4 0.2 0.833 189.7411 2. 400 2.5 0.17 0.833 1062.935 3. 600 4 0.2 0.833 389.2137 4. 117 2.5 0.275 0.833 -399.19 5. 400 4.6 0.275 0.833 -404.308 6. 683 2.5 0.275 0.833 -451.19 7. 200 1 0.2 0.833 -233.206 8. 600 1 0.2 0.833 150.9871 9. 400 2.5 0.275 0.833 -1456.21 10. 600 4 0.35 0.833 -772.317 11. 400 0.38 0.275 0.833 -525.381 12. 600 1 0.35 0.833 -693.28 13. 200 4 0.35 0.833 -520.967 14. 400 2.5 0.38 0.833 0 15. 200 1 0.35 0.833 -296.439 16. 400 2.5 0.275 0.833 0 17. 400 2.5 0.275 0.833 0 18. 400 2.5 0.275 0.8325 0 Table 5, The )*+ network weights matrices and bias. no. MN ON MP 1. 400 2.5 0.17 0.8326 0 2. 200 1 0.35 0.8326 15.3151 3. 200 4 0.35 0.8326 1.05552 4. 600 1 0.35 0.8326 3.96737 5. 600 4 0.2 0.8326 24.3887 6. 400 4.6 0.275 0.8326 0.59432 7. 683 2.5 0.275 0.8326 0.53165 8. 117 2.5 0.275 0.8326 0.49165 9. 400 0.38 0.275 0.8326 0.28587 10. 600 4 0.35 0.8326 -24.091 11. 200 1 0.2 0.8326 -14.408 12. 600 1 0.2 0.8326 -3.2356 13. 400 2.5 0.275 0.8326 2.88848 14. 400 2.5 0.38 0.8326 -2.3251 15. 200 4 0.2 0.8326 -0.1788 16. 400 2.5 0.275 0.8326 0 17. 400 2.5 0.275 0.8326 0 18. 400 2.5 0.275 0.8326 0 These networks are used to optimize the cutting parameters that produce lowest machining force and chip thickness ratio. To do that, a Matlab function has been written. The function creates two three-dimension arrays with (60, 60, 60) in size and initialized with zeros. The first array stores the results of �� and the second stores ���. In addition, its creates three vectors using the linspace MATLAB function. Each vector has 60 elements for the three parameters cutting speed, depth of cut and feed rate. The range of the cutting speed vector is (117 – 683) m/min, for depth of cut is (0.38 – 4.62) and for feed rate is (0.170 – 0.380). Then it performs loops to execute the networks with different parameters. The optimum parameters are founded by searching for minimum �� and minimum ��� arrayes. Table 6 presents the optimum parameters that gives lowest �� and ���. The optimum parameter for both the �� and ��� are differ only in the feed. A surface drawn when taking the optimum cutting speed value as constant and changing the remaining for both the computed �� from the networks output and ��� are shown in Fig. 9 and Fig. 10, respectively. The square correlation coefficients of the proposed models are compared to those given by Santos, et al. [5] as shown in Table 7. Table 6, The optimum parameters and their corresponding optimum values from &Q and )*+ �� R ��� S !���� " � �� #$% � Optimu m value �� 683 3.18 0.27 240.46 N ��� 683 3.18 0.17 1.21 Table 7, RBF neural network TP versus TP from [5]. +P "U# &Q +P "U# )*+ Proposed RBF networks 1 1 Results from Santos, et al. [5] 0.998 0.9661 Fig. 4 The neural network for &� response against the experimental data from Santos, et al. [5]. Mohanned Mohammed H. Al-Khwarizmi Engineering Journal, Vol. 14, No. 1, P.P. 67- 76 (2018) 73 Fig. 5 The neural network for &( response against the experimental data from Santos, et al. [5]. Fig. 6. The neural network for &" response against the experimental data from Santos, et al. [5]. Fig. 7. The neural network for CTR response against the experimental data from Santos, et al. [5]. Fig. 8. The &Q computed from the networks outputs against the experimental data from Santos, et al. [5]. Fig. 9. &Q surface when cutting speed is (683 m/min) Fig. 10. )+* surface when the cutting speed is (683 m/min). Mohanned Mohammed H. Al-Khwarizmi Engineering Journal, Vol. 14, No. 1, P.P. 67- 76 (2018) 74 5. Conclusion This study provided an experimental investigation, via radial basis function RBF network modeling, to estimate the affect of cutting parameters. (cutting speed, depth of cut and feed rate) on machining force (Fu) and chip thickness ratio (CTR) during turning of high strength aluminum alloy 7075-T6. The primary conclusions of the investigation are given following: 1. The proposed RBF networks showed an extreme match to the experimental data and the computed correlation coefficients were equal one. additionally, those networks were used to optimize the cutting process and obtain the optimum cutting parameters. 2. The proposed methodology based on RBF neural network modeling can effectively overcome any complicated function approximation with more than two inputs. 3. The outcome also revealed that the effectiveness of the developed networks was better compared to existing using genetic algorithm (GA). The present study for optimizing the cutting process is anticipated to open two directions which can be suggested to continue this work. The first is to investigate the effect of more cutting parameters, which include cooling liquids and angles of cutting tools, on machining force and chip thickness ratio. The second possible direction is to integrate a neural network with fuzzy logic to solve a more complicated function approximation models. Notation Depth of cut 71 Network first layer bias vector 7> Network second layer bias vector ��� Chip thickness ratio Feed rate �� Cutting force �� Passive force �� Feed force �� Machining Force 3 Network input vector � correlation coefficient �� Cutting speed 51 Network first layer weights matrix 51 Network first layer weights matrix 6. References [1] V. P. Astakhov and X. Xiao, "A Methodology for Practical Cutting Force Evaluation Based on the Energy Spent in the Cutting System," Machining Science and Technology, vol. 12, no. 3, pp. 325-347, 2008. [2] W. Stachurski, S. Midera and B. Kruszy, "Determination of mathematical formulae for the cutting force FC during the turning of C45 steel," Mechanics and Mechanical Engineering, vol. 16, no. 2, pp. 73-79, 2012. [3] B. d. Agustina, C. Bernal, A. Camacho and E. Rubio, "Experimental Analysis of the Cutting Forces Obtained in Dry Turning Processes of UNS A97075 Aluminium Alloys," Procedia Engineering, vol. 2004, no. 63, pp. 694-699, 2013. [4] C.X.Yue, X.L.Liu, D.K.Jia, S. c and Y.S.Zhai, "3D Finite Element Simulation of Hard Turning," Advanced Materials Research , Vols. 69-70 , pp. 11-15, 2009. [5] M. C. Santos, J. A. R. Machado, M. A. S. Barrozo, M. J. Jackson and E. O. Ezugwu, "Multi-objective optimization of cutting conditions when turning aluminum alloys (1350-O and 7075-T6 grades) using genetic algorithm," Machining with Nanomaterials: Second Edition, vol. 76, no. 5-8, pp. 323-346, 2015. [6] V. P. Astakhov and S. Shvets, "The assessment of plastic deformation in metal cutting," Journal of Materials Processing Technology , vol. 146, no. 2, pp. 193-202, 2004. [7] M. Chandrasekaran, M. Muralidhar, C. M. Krishna and U. S. Dixit, "Application of soft computing techniques in machining performance prediction and optimization: A literature review," The International Journal of Advanced Manufacturing Technology, vol. 46, p. 445–464 , 2010. [8] B. Sick, "Online tool wear monitoring in turning using time-delay neural\nnetworks," Neural computing and application, vol. 7, pp. 356-366, 1998. [9] V. S. Sharma, S. Dhiman, R. Sehgal and S. K. Sharma, "Estimation of cutting forces and surface roughness for hard turning using neural networks," Journal of Intelligent Manufacturing, vol. 19, no. 4, pp. 473-483, 2008. Mohanned Mohammed H. Al-Khwarizmi Engineering Journal, Vol. 14, No. 1, P.P. 67- 76 (2018) 75 [10] Y. Chen, R. Sun, Y. Gao and J. Leopold, "A nested-ANN prediction model for surface roughness considering the effects of cutting forces and tool vibrations," Measurement, vol. 98, p. 25–34, 2017. [11] K. S. Sangwana, S. Saxenaa and G. Kanta, "Optimization of Machining Parameters to Minimize Surface Roughness using Integrated ANN-GA Approach," in The 22nd CIRP conference on Life Cycle Engineering , 2015. [12] M. M. H. AL-Khafaji, H. L. Alwan and B. M. H. Albaghdadi, "Roughness Assessment for Machined Surfaces in Turning Operation Using Neural Network," Engineering and Technology, vol. 32, no. 5, pp. 1331-1344, 2014. [13] M. Mia and N. R. Dhar, "Prediction of surface roughness in hard turning under high pressure coolant using Artificial Neural Network," Measurement, vol. 92, pp. 464-474, 2016. [14] P. S. Pai, T. N. Nagabhushana and P. K. R. Rao, "Flank Wear Estimation in Face Milling Based on Radial Basis Function Neural Networks," The International Journal of Advanced Manufacturing Technology, vol. 20, pp. 214-247, 2002. [15] F. J. Pontes, A. P. d. Paiva, P. P. Balestrassi and J. R. Ferreira, "Optimization of Radial Basis Function neural network employed for prediction of surface roughness in hard turning process using Taguchi’s orthogonal arrays," Expert Systems with Applications journal, vol. 39, pp. 7776-7787, 2012. [16] H. Demuth, M. Beale and M. Hagan, Neural Network Toolbox User's Guide, Math Works, Inc., 2009. [17] M. T. Hagan, H. B. Demuth, M. H. Beale and O. D. Jesús, Neural network design, 2nd Edition ed., Oklahoma: Oklahoma State University, 2014. )2018( 67-76، صفحة1د، العد14دجلة الخوارزمي الهندسية المجلم مهند محمد حسين 76 نمذجة قوة القطع ونسبة سمك النحاتة باستخدام الشبكات العصبية اثناء خراطة سبيكة األلمنيوم 7075-T6 مهند محمد حسين الخفاجي العراق /بغداد/ / الجامعة التكنولوجيةقسم هندسة اإلنتاج والمعادن mohannedalkhafaji@hotmail.com البريد االلكتروني: الخالصة القطع واستهالك الطاقة ودرجة ةتؤثر الكثير من المتغيرات على عملية القطع والتي يجب دراستها. ومن هذه المتغيرات الخشونة السطحية وعمر عد معقدة وال خطية بسبب هذه العوامل. إن الهدف من هذا البحث هو دحرارة القطع ومركبات قوى التشغيل وتآكل العدة ونسبة سمك النحاتة. ان عملية القطع تع ة سمك النحاتة. ببناء نماذج من الشبكات العصبية لتمثيل العالقة بين متغيرات القطع (سرعة القطع وعمق القطع ومعدل التغذية) وقوة التشغيل وكذلك مع نس ذلك تم فضالعنث شبكات عصبية نصف قطرية لكٍل من قوة القطع والقوة السلبية وقوة التغذية، . تم بناء ثالT6-7075عملة الخراطة اجرية لسبيكة األلمنيوم ءمعدل التغذية). تم مقارنة أداالقطع وانشاء شبكة نصف قطرية لنمذجة نسبة سمك النحاتة. ان مدخالت جميع الشبكات هي ظروف القطع (سرعة القطع وعمق لعالقة وكذلك تم حساب معامل ا اتام اتطابقمع التجارب العملية وأعطت غيل (قوة القطع والقوة السلبية وقوة التغذية) لمركبات قوة التش الشبكات (مخرجات) بكات شمعامل االرتباط مساٍو للواحد ايضاً، بالمقارنة مع النتائج العملية. ان هذه الها لنسبة سمك النحاتة كان ؤتم بناوكذلك الشبكة التي وجد بأنه مساٍو للواحد. أظهرت نتائج النماذج بإن افضل قوة تشغيل يمكن التي بدورها تعطي اقل قوة قطع وأقل نسبة سمك النحاتة.وظروف قطع أفضلاستخدمت إليجاد )نماذج(ال ملم/دورة). وأظهرت الشبكة ٠٫٢٧ملم) و معدل تغذية (٣٫١٨م/د) و عمق القطع ( ٦٨٣نيوتن) عندما تكون سرعة القطع ( ٢٤٠٫٤٦الحصول عليها هي ( ملم) و معدل تغذية ٣٫١٨م/د) وعمق قطع ( ٦٨٣عة القطع () عندما تكون سر١٫٢١اقل نسبة يمكن الحصول عليها هي (إن المقترحة لنسبة سمك النحاتة .ملم /دورة) ٠٫١٧(