Sie sind auf Seite 1von 9

Int. J.

Mach, Tools

Manufact.

Pergamon
0~0-.69SS(95)000194

Vol. 35. No. 12. pp. 1693-170|. 1995 CopyrightC) 1995ElsevierScienceLtd Printed in Great Britain. All rishts reserved 089(~3955/95S9.5(I + .(10

DETERMINATION OF OPTIMAL CUTYING PARAMETERS IN WIRE ELECTRICAL DISCHARGE MACHINING


Y. S. TARNG, S. C. MA and L. K. CHUNG
(Received 4 October 1994)

Al~r~:t---Owing to the complexity of wire electrical discharge machining (wire-EDM), it is very difficult to determine optimal cutting parameters for improving cutting performance. The paper utilizes a feedforward neural network to associate the cutting parameters with the cutting performance. A simulated annealing (SA) algorithm is then applied to the neural network for solving the optimal cutting parameters based on a performance index within the allowable working conditions. Experimental results have shown that the cutting performance of wire-EDM can be greatly enhanced using this new approach.

1. INTRODUCTION

In recent years, wire-EDM has become an important nontraditional machining process, widely used in the aerospace and automotive industries [ 1]. This is because wire-EDM provides an effective solution for machining hard materials with intricate shapes. However, selection of cutting parameters for obtaining higher cutting efficiency or accuracy in wire-EDM is still not fully solved, even with the most up-to-date CNC wireEDM machine. This is mainly due to the complicated stochastic process mechanisms in wire-EDM [2]. As a result, the relationships between the cutting parameters and cutting performance are hard to model accurately. Neural networks are highly flexible modeling tools with an ability to learn the mapping between input variables and output feature spaces [3]. The superiority of using neural networks in modeling machining processes has been reported in several references [4, 5, 6]. Therefore, neural networks are considered in this paper to model the wire-EDM process with large dimensional input and output spaces. On the basis of the developed neural network model, the effect of the input of cutting parameters (pulse on time, pulse off time, peak current setting, no-load voltage, servo reference voltage, capacitor setting, servo speed setting, and workpiece thickness) on the output of cutting performance (machining speed and surface finish) can be obtained. Once the process model of wire-EDM is constructed, an appropriate optimization algorithm with a performance index (objective function) is then carried out for searching the optimal cutting parameters. In this paper, a sound optimization method of simulated annealing (SA) [7] is adopted. Traditionally, the annealing process, used in metal working, involves heating the metal to a high temperature and then letting it gradually cool down to reach a minimum energy state. The SA algorithm introduced by Kirkpatrick, Gelatt, and Vecchi in 1983 [7] is a simulation of the annealing process for minimizing the performance index. It has been shown that SA can provide an effective way to jump out from a local optimum and to draw near global optimum [8]. As a result, the SA algorithm has emerged as a general tool for optimization of arbitrary functions with large dimensional spaces and has been successfully applied in VLSI layout generation [9], noise filtering in image processing [10], discrete tolerance design [11], etc. In the rest of the paper, the input and output variables of the wire-EDM process are given first. The experimental details of using a neural network in modeling of the

Department of Mechanical Engineering, National Taiwan Institute of Technology, Taipei, Taiwan, 106, R.O.C. 1693

1694

Y.S. Tarng et al.

wire-EDM process are then described. The theory of SA is briefly introduced and applied to searching the optimal cutting parameters in wire-EDM. Finally, an experimental verification of the optimal cutting parameters is presented.
2. DESCRIPTION OF THE WIRE-EDM PROCESS

Wire-EDM is an advanced material removal process using thin copper or brass wire as the tool electrode (Fig. 1). The workpiece and electrode are separated by deionized water as a dielectric medium. The wire travels through the workpiece in a closely controlled manner for generating spark discharges and then erodes the workpiece to produce complex shapes based on the paths of the tool electrode. During the last decade, computer numerically controlled (CNC) wire-EDM has become the major development in the wire-EDM technologies. In this paper, the experiments were performed on a Sodick A350-S CNC wire-EDM machine. The wire tension and speed are automatically adjusted by the wire-EDM machine to avoid wire breakage during machining. On the basis of the machining condition table [12], several cutting parameters which control the pulse energy and cutting speed (Fig. 2) need to be determined by operators. These cutting parameters can be set in a range of steps expressed as two digits, that is:
TI:

Ip:
Vo:
SV: C: SF:

pulse on time [00-09], pulse off time [00-31], peak current setting [00-17], no-load voltage [00-03], servo reference voltage [00-09], capacitor setting [00-03], servo speed setting [00-09].

In this paper, the cutting performance of wire-EDM is evaluated by the two output variables, that is:
Rrnax :

FR:

machined surface roughness (ixm), machining speed (mm min-1).

~:~ Wire

"-~

Discharge Spark

Discharge Mark Machined Particles e Fig. 1. Schematic diagram of the wire-EDM process.

Optimal Cutting Parameters

1695

a)
V
I- ~'w "1

T~

T~

T(time)

T(time)
Fig. 2. Typical discharge pulse wave forms: (a) voltage; (b) current. 3. NEURAL NETWORK MODELING OF THE WIRE-EDM PROCESS

A feedforward neural network is adopted here to model the wire-EDM process. The feedforward neural network is composed of many interconnected artificial neurons that are often grouped into input, hidden and output layers. The artificial neuron evaluates the inputs and determines the strength of each one through its weighting factor calculated by the back-propagation learning algorithm [13]. The weighted inputs are summed to determine the output of the neuron using a sigmoidal transfer function. The output of the neuron is then transmitted along the weighted outgoing connections to serve as an input to subsequent neurons. In this study, the neurons of the input and output layers are used to receive the input variable of cutting parameters and to send out the output variable of cutting performance, respectively. The workpiece thickness (H) is also considered as an input variable. As a result, there are eight input variables and two output variables in the neural network. To properly map the input and output relationships in the wire-EDM process with the neural network, finite discrete samples of experimental data are required for training the neural network. SUS-304 stainless steel specimens with different thicknesses were used in the experiments. A 28-4 fractional factorial experiment was designed to reduce the number of training samples [14]. The machined surface roughness (Rmax) was measured by a 3D-Hommelewerk profilometer. During the training process, the number of neurons in the hidden layer is determined by trial-and-error experimentation. It is found that a single hidden layer with 14 neurons can provide better convergence in modeling the wire-EDM process. As shown in Fig. 3, the roots mean square (RMS) error between the desired and predicted outputs is almost reduced to zero after 5000 iterations during the training process. Therefore, a feedforward neural network with an 8-14-2 type (Fig. 4) is adopted here to associate the cutting parameters with the cutting performance. In the following, the effect of the cutting parameters on the cutting performance will be studied one by one based on this developed neural network. In reality, cutting parameters affect the cutting performance of one another. To separate the effect caused by each cutting parameter, the other cutting parameters are set to a middle value in the allowable working spaces when one of the cutting parameters is varied and analyzed. Figures 5 and 6 show the effect of the variations of cutting parameters on the machining speed and machined surface roughness, respectively. An explanation of the effect of

1696 1.0 0.8

Y.S. Tarng et

al.

o 0.6
(D

0.4

0.0

J '~-'~L
0 2000 4000

6000

8000

Number of Iterations Fig. 3. Root mean square error vs number of iterations in the training process.

Tl T2 lp Vo SV C SF II Machining Speed Surface Roughness

Fig. 4. Configurationof the neural network. each cutting parameter on the machining speed and machined surface roughness is given as follows. (1) Pulse on time (T1): It is shown that machining speed increases with the pulse on time. On the contrary, surface accuracy decreases with increasing the pulse on time. This is because the discharge energy increases with the pulse on time. As a result, machining speed becomes faster with the increase of the discharge energy. However, in the meantime, the discharge gap becomes wider so as to decrease surface accuracy. (2) Pulse off time (T2): As the pulse off time is shorter, the number of discharges within a given period becomes more. This will lead to a higher machining speed. But, surface accuracy becomes poor because of a larger number of discharges. (3) Peak current setting (lp): The higher the peak current setting, the larger the

Optimal CuRing Parameters

1697

1.2

Ip 1.0 ~ 0.8 T1

Vo
SF

0.6 0.4

0.2 0.0 -2

Ic

i 0 2 4 6 8 10 12 14 16 18 20 22 24 26 28

30

Number of Steps Fig. 5. Effectof the cuttingparameterson the machiningspeed.


16.0

T1
14.0

)P

12.0

lO.O
V
8.o

6.0 -2

i i i i i i i ~ i j i ~ 0 2 4 6 8 10 12

~ i ~ i l'li 14 16 18

i ~ i ~ i ~ i ~ i ~ 20 22 24 26 28 30

Number of Steps Fig. 6. Effectof the cuttingparameterson the surfaceroughness. discharge energy. In other words, the effects of the peak current setting and pulse on time on the cutting performance are similar. But, the sensitivity of the peak current setting on the cutting performance is stronger than that of the pulse on time. When the peak current setting is too high, wire breakage may occur frequently. (4) No-load voltage (Vo): Basically, the no-load voltage is proportional to the peak current value according to the Ohm law. Therefore, the discharge energy increases with not only the peak current value but also the no-load voltage. As mentioned before, machining speed becomes quicker but surface accuracy becomes poorer owing to an increase of the discharge energy. (5) Servo reference voltage (SV): As shown in Fig. 7, the higher the servo reference voltage, the longer the discharge walt time ('rw). To obtain the longer discharge walt time, the machining speed needs to be slowed down. This will lead to a wider average discharge gap. Therefore, the discharge condition becomes more stable but the number

1698

Y . S . Tarng et al.

a)
Z'Wl

I"

~v I

--1

--1
T(time)

b)
I- -I

~'W2

--I

--I
T(time)

Fig. 7. Effect of the servo reference voltage on the discharge wait time ('rwa > 7w2 and SV~ > SV2).

of discharge cycles decreases within a given period. Owing to this stable machining, surface accuracy becomes better. (6) Capacitor setting (C): The capacitor setting is used to determine whether or not it is necessary to insert a capacitor circuit between the workpiece and electrode. Usually, the use of the capacitor circuit can generate a larger shock applied to the discharge energy. It will help the discharge condition to be more stable. As shown in Figs 5 and 6, the effect of the capacitor setting on the cutting performance is not apparent in comparison with the other cutting parameters. (7) Servo speed setting (SF): The servo speed setting is used to set the table feed speed when there is no load across the workpiece and electrode or the average voltage is higher than the servo reference voltage (SV). For the machine used in the experiments, a large setting value corresponds to a small table feed speed. Generally, the discharge gap becomes narrower when the larger table feed speed is selected. As a result, the machining speed increases with the table feed speed. However, surface accuracy becomes poor due to the fast machining speed. In the previous discussion, the complicated relationships between the cutting parameters and cutting performance are clearly clarified using the developed neural network. The results obtained from the neural network are consistent with the machining condition table [12], provided by the builder of the CNC wire-EDM machine. Therefore, the developed neural network is a suitable tool for modeling of the wire-EDM process.
4. S I M U L A T E D A N N E A L I N G A L G O R I T H M

In condensed matter physics, annealing is referred to as a physical process that is used to reconstruct the crystal structure of a solid with a low energy state. A solid in a heat bath is first heated up to a temperature above the melting point of the solid. At this temperature, all particles of the solid are in violent random motion. The temperature of the heat bath is then slowly cooled down. All particles of the solid rearrange themselves and tend toward a low energy state. As the cooling of the particle is carried out sufficiently slowly, lower and lower energy states are obtained until the lowest energy state is reached. In this annealing process, the probability Pr(E) of being in an energy state E can be expressed by the Boltzmann distribution, that is:

Optimal Cutting Parameters


= I

1699

P,(E)

Z~

exp(- k~)

(1)

where Z(T) is a normalization factor, depending on the temperature T, and k , is the Boltzmann constant. As shown in eqn (1), the probability P,(E) approaches to one for all energy states when the temperature T is high. However, as the temperature T drops, the probability of high energy states decreases as compared to the probability of low energy states. It implies that the Boltzmann distribution concentrates on the low energy states as the temperature decreases. In 1953, Metropolis et al. [15] proposed an algorithm that is used to simulate the cooling of a solid for reaching a new energy state. Given the current energy state of the solid with the temperature T, a small randomly generated perturbation is then applied to the particles of the solid for generating a perturbed energy state. If the energy difference AE between the current and perturbed states is negative, the particles of the solid rearrange themselves and tend toward a low energy state. Therefore, the perturbed energy state is accepted as the new energy state. However, if AE is positive, the probability of the acceptance of the perturbed state is expressed as:

P,(E) = e x p ( - k~T)

(2)

This acceptance rule for the new state is called the Metropolis criterion. After about 30 years, Kirkpatrick et al. [7] were starting to use the Metropolis criterion to minimize the objective function in a combinational optimization problem and developed an optimization algorithm called "simulated annealing". Basically, the simulated annealing algorithm can be treated as a sequence of Metropolis algorithms evaluated by a sequence of decreasing temperatures. The objective function that needs to be minimized in an o p t i ~ a t i o n problem is equal to the energy function used in the Metropolis criterion. In recent years, the simulated annealing (SA) algorithm has emerged as a leading tool for large-scale combinational optimization problems and successful applications of simulated annealing to engineering optimization have been reported continuously [16]. Figure 8 shows the graphical representation in the SA search of the gobal minimum
( ) Box

Objective F~vtctiovt

f(x)
~- C u r r e n t Solution

/
Local M%~%rr~um

\ /

Local

Global Mi~w~z~

Solutio'r~ (X)
( }

Fig. 8. Graphical illustration of the simulated annealing search of the gobal minimum.

HIH 3S-12-6

1700

Y.s. Tarng et al.

of an objective function. It is shown that a marble moves on a surface in a box. When the temperature is high, the box is shaken violently in the horizontal direction. The marble then moves up and down rapidly and never settles at a fixed point. If the shaking is strong enough, the marble can climb up the hill so as to jump out of the areas of the local minimum. As the temperature decreases, the violence of the shaking is gradually reduced. Finally, the marble will stay at the point of the gobal minimum. In the foregoing discussion, the SA algorithm processes several advantages in comparison with the traditional optimization algorithms. First, the SA algorithm does not need to calculate the gradient descent that is required for most traditional optimization algorithms. This means that the SA algorithm can be applied to all kinds of objective and constraint functions. Next, the SA algorithm can find the gobal minimum more efficiently instead of trapping in a local minimum where the objective function has surrounding barriers. Furthermore, the SA search is independent of initial conditions. Because of these attractive properties, the SA algorithm is employed in this paper to solve the optimal cutting conditions in wire-EDM. 5. OPTIMAL CUTTING PARAMETERS IN WIRE-EDM To determine the optimal cutting parameters, the objective function has to be defined first. As shown in Figs 5 and 6, the machining speed and machined surface roughness are the two objective functions that are required to be optimized or minimized. A simple weighting method is used to transform the two objectives into a single objective format. Thus, the objective function ( P I ) can be expressed as:
P1 = W1 Rmax - W2 F R

(3)

where Rrnax and F R are the normalized surface roughness and normalized machining speed in the range of [0, 1]; W1 and W2 are the weights of the normalized surface roughness and normalized machining speed. The optimal cutting parameters can be obtained by an adjustment of the weights. For instance, a fine surface roughness is the major concern in finish machining. The corresponding optimal cutting parameters can be obtained by increasing the value of W1. On the other hand, the optimal cutting parameters corresponding to a high machining efficiency in rough machining can be found by increasing the value of WE. In the following, two rough machining examples are presented to illustrate the approach developed in this study. The values of W1 and W2 are equal to one and ten, respectively. In the first case, the workpiece thickness is 10 mm. The setting of the seven cutting parameters obtained from the optimization technique is listed in Table 1. As indicated in Table 1, the error between the predicted and actual cutting performances is reasonably small. In the second case, the workpiece thickness is changed to be 15 mm. It is shown that the actual cutting performance is still very close to the predicted cutting performance (Table 2). Table 1. Optimization results for case 1 Thickness 10 mm
T~ T2 lp Vo SV C SF

27

17

Optimal value Prediction Experiment Error (%)

Rm.x (p,m) 16.1 16.3 1.2

FR(mm/min)

1.63 1.74 6.3

Optimal Cutting Parameters Table 2. Optimization results for case 2 Thickness 15 nun
7"1 T2 lp Vo SV C SF

1701

15

17

Optimal value Prediction Experiment Error (%)

R=.~ (pan) 17.1 17.3 1.2

FR (mm rain -1)

1.65 1.70 2.9

6. CONCLUSIONS In this p a p e r , a f e e d f o r w a r d neural n e t w o r k is used to construct the w i r e - E D M process model. It is shown that the neural n e t w o r k can clearly clarify the complicated relationships b e t w e e n the cutting p a r a m e t e r s and cutting p e r f o r m a n c e . O n c e the process m o d e l is established by the neural n e t w o r k , a gobal optimization algorithm, simulated annealing ( S A ) , is then applied to this n e t w o r k for solving the optimal cutting param e t e r s using an adjustable objective function. E x p e r i m e n t a l results have shown that the use of the a p p r o a c h can systematically search the cutting p a r a m e t e r s for obtaining an o p t i m u m p r o d u c t i o n rate in w i r e - E D M . T h e r e f o r e , the efficiency of determining optimal cutting p a r a m e t e r s in the process planning of w i r e - E D M can be dramatically i m p r o v e d by using this a p p r o a c h . 7. REFERENCES [1] J. A. McGeough, Advanced Methods of Machining, Chapman and Hall, New York (1988). [2] K. P. Rajurkar, Technology and research in EDM and ECM, In ASME Fundamental Issues Mach., 43, 309-336 (1990). [3] J. A. Freeman and D. M. Skapura, Neural Networks: Algorithms, Application, and Programming Techniques, Addison-Wesley, New York (1991). [4] S. Rangwala and D. Dornfeld, Learning and optimization of machining operations using computing abilities of neural networks, IEEE Trans. Syst., Man, Cybernetics, 19(2), 299-314 (1989). [5] T. W. Liao and L. J. Chert, A neural network approach for grinding processes: modeling and optimization, Int. J. Mach. Tools Manufact., 34(7), 919-937 (1994). [6] B. Y. Lee, Y. S. Tarng and S. C. Ma, Modeling of the process damping force in chatter vibration, Int. J. Mach. Tools Manufact., 35(7), 963-974 (1995). [7] S. Kirkpatrick, C. D. Gelatt and M. P. Vecchi, Optimization by simulated annealing, Science, 220(4958), 671-680 (1983). [8] R. A. Rutenbar, Simulated annealing algorithms: an overview, IEEE Circuits and Devices Magazine, 5(1), 19-26 (1989). [9] B. W. Lee and B. J. Sheu, Hardware Annealing in Analog VLSI Neurocomputing, Kluwer Academic Publishers, London (1991). [10] S. Geman and D. Geman, Stochastic relaxation, Gibbs distributions and the Bayesian restoration of images. IEEE Trans. on Pattern Analysis and Mach. Intelligence, 6(6), 721-741 (1984). [11] C. Zhang and H. P. Wang, The discrete tolerance optimization problem, ASME manufacturing review, 6(1), 60-71 (1991). [12] Sodick Co., Ltd, Fine sodick wire cut-EDM A350-S machining condition table, Vol. 1.2G, Yokohama, Japan (1992). [ 13] D. Rumelhart and J. McCelland, Parallel Distributed Processing, Vol. 1, MIT Press, Cambridge (1989). [14] D. C. Montgomery, Design and Analysis of Experiments, Wiley, Singapore (1991). [15] N. Metropolis, A. Rosenbluth, M. Rosenbluth, A. Teller and E. Teller, Equation of state calculation by fast computing machines, J. of Chem. Physics, 21, 1087-1092 (1953). [16] P. J. M. van Laarhoven and E. H. L. Aarts, Simulated Annealing: Theory and Applications, Kluwer Academic Publishers, London (1988).

Das könnte Ihnen auch gefallen