ARCHIVES of International Scientific Journal

Computational published quarterly by the Association

Materials Science of Computational Materials Science
and Surface Engineering and Surface Engineering

2 0 0 9 • Vo l u m e 1 • Is s ue 3 • 161- 167

EDM process optimisation via predicting

a controller model
R.A. Mahdavinejad*
School of Mechanical Engineering, Faculty of Engineering, University of Tehran, Tehran, Iran
* Corresponding author: E-mail address: mahdavin@ut.ac.ir
Received in a revised form 06.05.2008

ABSTRACT

Purpose: Electro-discharge machining is an important manufacture technology in machining difficult-to-cut materials

and to shape complicated contours and profiles with high material removal rate, low tool wear and good tolerances.
Design/methodology/approach: In machining of carbon-based materials such as WC-Co and non-oxide ceramics
which are growingly used, the complexity and non-linear nature of EDM is a serious problem. EDM is the best and
nearly the only non-conventional method for machining of these kind of materials, but it shows high instability and
tendency to arcing, compared with machining of steel. Occurrence of instability phenomenon due to the different
input setting up parameters make the modeling of EDM process impossible with conventional methods. To achieve
instantaneous data from machining condition, the new method of fuzzy analysis of single machining pulses and
computing the magnitude of system condition in the form of a real number between 0 and 1, has been used.
Findings: Some tests with WC-Co material are carried out and finally, the results of implementation of control
system on a sinking ED machine and an EDM system that has been set with an expert user, has been compared.
Practical implications: The optimization and control of EDM process using the neural model predictive control
method. A genetic algorithm has also been employed to optimize the input parameters and to create the optimized
setting collection of process.
Originality/value: The testing results from ED machining of WC-Co confirms the capability of the system of
predictive controller model based on neural network with 32.8% efficiency increasing in stock removal rate.
Keywords: EDM; Optimisation; Fuzzy logic; Model predictive control
Reference to this paper should be given in the following way:
R.A. Mahdavinejad, EDM process optimisation via predicting a controller model, Archives of Computational
Materials Science and Surface Engineering 1/3 (2009) 161-167.

MANUFACTURING AND PROCESSING OF ENGINEERING MATERIALS

1. Introduction
1. 
Introduction magazine, tools magazine, workpiece magazine, laser zero point
measurement system, cutting forces dynamometers for process
diagnostics, frequencies sensor, acoustic sensors [8]. The
The object of special interest of researchers is increasing the
instability phenomenon in electro discharge machining (EDM) is
machine’s efficiency and reliability, during the process of
the most important factor in rapid tool wear and great decrement
designing the machines [1]. The necessity for frequent product
in stock removal rate. Occurrence of this phenomenon especially
line changes causes the development in flexible production
in machining of carbon-based materials such as carbides and non-
systems [9]. Some specifics of modern machine tool are included
oxide ceramics due to carbon immanency of workpiece and
over modules, as rotation table, precise high RPM spindle, bar

© Copyright by International OCSCO World Press. All rights reserved. 2009 161
 R.A. Mahdavinejad

convenient dielectric is inescapable. The researches show that the controller, a genetic algorithm on parameters of pulse on and off
arcing phenomenon as the symbol of instability in EDM, has time, discharge current, gap size and its variations rate, are used.
unknown stochastic distributions caused by different input Finally, the method of execution of changes on a convenient ED
parameters. According to instantaneous changes in discharge machine is explained.
conditions such as dielectric pollution, the physical model for the
process is valid only for small time durations which cause the
instantaneously changes in arc distribution function factors. Input
setting parameters to obtain the desired outputs such as high stock
2. EDM
2.  Instabilityprocess
EDM instability Process
removal rate and low tool wear ratio has been accomplished
In electro discharge machining process, to create plasma
manually based on user’s experience and without any
channel and efficient discharges, a high potential difference is
optimization. Generally, methods used to model the processes in
used between tool and workpiece. The electrical field between
form of “black box” has been suggested for modeling of such
two electrodes emerged in dielectric tank causes ionization, ion
processes. Therefore, a method for physically modeling will be
stimulation and creation of vapor and plasma channel. The plasma
effective and reliable that updates its knowledge about the process
channel will be out of concentration and an approximately
regularly. One of the most fundamental criteria in the design of
constant current with low oscillation will pass through the gap
modern mechanical structures are their dynamical propreties, as
whenever the necessary condition of discharge is provided. In
they have a direct impact on the vibrations of the system, noise
most EDM applications, especially in die sinking EDM, the
emission, fatigue resistance, controllability and stability [6].
hydrocarbon liquid with linear structure, is used as dielectric. In
Challenging problems for scientific research are the requirements
machining of carbide materials, the existence of free carbon
concerning mechatronic system, for example their exact
created from dielectric decomposition with the high chemical
positioning working velocity, control and dimensions [2].
activity from one hand, and the immanency between this free
Optimization of non-linear models with unknown transfer
carbon and that one in the structure of carbide workpiece on the
functions usually is possible using methods based on artificial
other hand, causes the carbon deposition on the workpiece and
intelligence. The first research to distinguish the pulse type with
afterwards, tendency to arcing. So that, the continuance of the
artificial intelligence is carried out in 1997 by Kao and Tarng [7].
process leads to form a carbon bridge between electrodes and
They tried to make distinctions among pulse types instantaneously
makes the machining impossible. Some other factors such as; gap
via artificial neural networks. Also, Liu and Tarng in the same
size not exactly set up and tool's head reaction speed, due to the
time modeled the EDM using neural networks. In 1997 Spedding
abnormal pulses, also have affection in creation of non-successive
and Wang tried to optimize the surface characterization and input
pulses as open circuits and short circuits. Generally inexact setting
parameters via modeling of wire-EDM process [10]. According to
of each input parameter makes the machining process unstable.
the determination of transformation function via the mathematical
Anyway, the pulse shape recorded from machining zone can be
relations between input and output vectors, the capability
used as an indicator to determine the pulse types. Variations of
computation of the fitness function value in optimization process
gap voltage, discharge current, discharge delay time and high
will be determined [11]. On the other hand, in the process not
frequency elements on voltage-time recorded diagrams can be
being mathematically formulated, the relationship between input
used to distinct normal from other pulse types. Up to now, the
and output vectors must be presented, so that the output vectors
distinction of pulse type has been done based on average gap
with any input and desired accuracy respect to the physical
voltage, due to its easy accessibility and controllability.
process, should be available [4]. The artificial neural networks are
According to the time constant measuring device, the low level of
the best choice to model the physical processes in the black box
gap voltage represents the large number of arc and short circuit
format [7,12]. Although parameters such as stock removal rate,
and its high level represents the large number of open circuit
tool wear ratio and surface roughness are the main output
pulses. The middle bound of gap voltage approximately means
parameters of EDM process; it is possible to achieve high stock
that the system is working normally. However, due to possibility
removal rate and low tool wear ratio with good surface finishing
of arcing at any setting of system, even a suitable level of gap
if the stability of machining process as an effective mode in
average voltage, means that there are probably all kinds of
output parameters is guaranteed [12]. There are four completely
discharge pulse types in pulses train. Significant changes in
distinct pulses in EDM and this system is unstable when the
average gap voltage from normal to higher or lower levels,
number of non-successive pulses such as; arcing, short circuit and
indicates the system instability. Control of average gap voltage to
open circuit are increasing against normal discharges during
maintain the effective gap size is the simplest method to control
machining process. Although the voltage frequency
the EDM process, but, it is not necessity the best way. The best
characteristics of normal discharge is completely different from
method for this subject is indeed the method based on multi-factor
non-successive pulses, the system's output covers wholly a
indecisive decision.
continuous space, so that, the stability and non-stability of the
system is not quite clear.
In this paper, the method of model predictive control based on
artificial neural networks with output parameters of the system to 3. Pulse
3.  Distinctionusing
Pulse distinction Usingfuzzy
Fuzzy Logic
logic
minimize the number of non-successive pulses is used. To
determine the value of stability process parameter, a fuzzy Up to now, the distinction of pulse type based on detection of
analysis method is developed to distinguish the single pulse normal discharge from arc and other abnormal pulses has been
discharge type. As the optimizing algorithm of model predictive accomplished with different methods. Determination of a

162 Archives of Computational Materials Science and Surface Engineering

EDM process optimisation via predicting a controller model

threshold on discharge voltage so that its lowest level is higher Therefore, this system acts as a fuzzy sensor for machining
than arcing voltage level, measuring the discharge delay time process. This subject is very important in closed loop control
which is zero or very small in quasi arcing and arcing discharge systems. Figure1 shows the number of pulses recorded from
and analysis of high frequency elements on gap voltage-time machining zone with open-circuit voltages equal to 250V. The
diagram pulse by pulse, are the fundamentals of classical pulse condition factor of the system is calculated from 15 sample of the
distinction methods in EDM. Although the pulse type distinction pulse train.
with artificial intelligence-based methods are usually successful,
but, the stability analysis of the system via studying of single
pulse and assigning a relative instability factor is possible only by
using fuzzy logic. The fuzzy approach to microscopic pulse
distinction obtained from pulse train in gap, is based on two main
reasons:
x The stable and unstable machining boundary zone especially
in carbide materials is quite ambiguous.
x Too many linguistic rules are used by EDM specialists in
pulse interpretation. The second expression not only shows
the fuzzy analytical ability of pulse train, but, the decision
based on fuzzy interpretation in EDM process [10]. An
example of such linguistic rules about single-pulse Fig. 1. Typical pulse train sampled by an oscilloscope
interpretation is; if<discharge delay time is small> and
<average gap voltage is low> then <the discharge is arc>.
x The factors used in fuzzy analysis are as follows: 4. 
Modeling of
4. Modeling of EDM
EDMProcess
Process
x Average discharge voltage from breakdown to the end of
pulse on-time.
x Discharge delay time. A neural network is a bulky system of interconnected
calculator elements which are connected parallelism to each of
x High-frequency elements on pulse voltage-time diagram.
these elements named node does not have capability of data
The third factor is neglected due to the unnecessary of process
acquisition separately, but, the necessary data to interpolate are
analysis in frequency field and avoiding of massive calculations.
stored along several connected nodes.
The first two mentioned factors determine the pulse type by using
In recent years, the artificial neural network is transformed to
linguistic rules bellow:
a very useful tool for modeling complex systems. This capability
x If <the delay time is too short> and <the average discharge
is mainly because of the high power of neural network in learning
voltage is low> then <the pulse is arc>.
and education. Besides, the behaviour of artificial neural networks
x If <the delay time is medium> and <average discharge is similar with predicting systems. Neural networks are able to
voltage is medium> then <the pulse is normal>. model non-linear processes via catching the desired input and
x If <the discharge average voltage is too low> then <the pulse output vectors and the training of homologous algorithm. In the
is short circuit>. other word, the training capability of neural networks eliminates
x If <the delay time is long> and <average discharge voltage is the needs of explicit equation for process modeling.
high> then<the pulse is open circuit>. In this research the three layered perceptions neural network
The above linguistic rules with fuzzy of two parameters and sigmoid transfer function at the first and second layer and
(discharge delay time and average discharge voltage) and suitable linear function at external layer, is used to model EDM sinking
fuzzy operators' creation and also using appropriate fuzzy related process. In 1987 Cybenco and Nielsen proved that this network is
to the input parameters will be able to allocate a number between able to estimate any non-linear function with desired accuracy [5].
0 and 1 to the discharge condition of any pulses. Therefore, after The neuron numbers of external layer is determined with the
nom-fuzzy operations, this fuzzy analysis is able to determine number of output parameters. The error back propagation is used
numerically the condition of any machining pulses. For n pulses as a training algorithm. This kind of neural network is very usual
recorded, the condition factor of system is as follows: and skilled in simultaneously monitoring and at the same rank,
1 n very simple in communications and network structure.
Ok ¦ J i , k 1,2,, n
ni1
(1)

where J k is numerical average condition factor of each pulse. 4.1. Multi-layer perceptions neural network
It is obvious that, the more number of pulses recorded, the higher
4.1. Multi-Layer Perceptions Neural Network
accuracy in condition factor of system results, indeed. This
The multi-layer perceptions neural network is formed from
subject refers to the situation of dielectric in machining zone, so
numerous neurons with parallel connection, which are jointed in
that, any debris in this position may cause abnormal discharges.
several layers [3]. The structure of this network contains of
Therefore, increasing in number of pulses recorded decreases the
network's input data, numbers of hidden middle layers with
total errors. In this manner, condition factor of the system
numerous neurons in each layer and an external layer with
especially after any set up variation in current domain, pulse on
neurons connected to output. A perception neural network with
and off time, gap size and the speed of head displacement due to
three layers as matrix form is shown in Figure 2.
voltage variation, will be determined via the fuzzy system created.

Volume 1 • Issue 3 • 2009 163

 R.A. Mahdavinejad

Fig. 2. Matrix demonstration of supposed neural network

Fig. 3. Diagram of Nero-model predictive control method

This kind of network due to its sigmoid transferred function in
the middle layers and linear transferred function at the external
layer has universal approximation capability. According to the
This model makes available the mathematical model for non-
complexity and reciprocal affections of various machining
linear and complexity systems with low information and different
parameters, the nature of EDM process is not completely cleared
behaviours. When the input and output parameters are also
to the user; therefore, the kind of the selected network should
numerous, neural network is considered as first solution. Feed
have universal approximation capability. Generally, the
forward Perceptron network as multi-layered neural network
application of linear function in the last layer, guarantees the
which has universal approximation conditions is used.
adhesion of output parameters. Input and output parameters
Constrained optimization can also be carried out via an arithmetic
classification defines the cell's kind of network somehow. In
explicit method or with a method based on artificial intelligence
particular, the single output parameter of the desired network is
as genetic algorithm. In optimization operation, it should be under
considered as EDM condition system. This parameter is the
consideration that the total converging time of optimization
output of fuzzy sensor and a real number between zero and one.
algorithm and training of neural network, determines the factor of
This output parameter is a direct function of the input of the
total starting time of the controller. But, during the control
network, gravimetric values and connected bias terms. Therefore,
process, the length of control cycle is usually low. Therefore, the
naturally to create a meaningful similarity between input and
starting of controller can be divided into three times phases as
output vectors terms in data training network, the components
follows:
value of gravimetric matrix and bias vector in each neuron
x Neural network training phase with real data.
connection, is varied. The network training with real data means
the creation of optimized values of components in gravimetric x Data phase optimization according to neural model.
matrixes and biases with the aim of one by one similarity creation x Up to dating phase of the model via data observed during
between input and output parameters terms. The error back control operation if it is needed.
propagation and Levenbergg-Marquardt algorithms are used for The third phase in process in which the physical conditions
training of multi-layered perception neural networks. On and off and the model between input and output are variable during
pulse times, discharge current, gap size and its variation rate are various process operations is very important. In this model,
the input parameters of network. Most of these parameters are generally, the primary model which is trained with numerous
varied steeply and non-continually in their domain and indeed the input and output couples and out of training data during
machines structure does not let them to vary continually. But, processing, are used to up to date neural model. This continuous
usually, the setup steeply gap, does not make so much error to training process causes to remain the network knowledge with
cause serious variation in the stage of operational sets. It is good accuracy in desirable level.
possible even to set up the step duration and afterwards error The control of predicting EDM model via neural network to
minimization via small software and hardware variations. optimize WC-Co machining process on a die sinking machine
with gap control system is carried out. The capability of the
system with and without gap control system is compared. Table 1
shows the testing conditions.
5. 
Control predicting
5. Control Predictingmodel
Model According to the criteria system condition the train of pulses
recorded, the system stability factor is defined as follows:
The control predicting model is not usually an exact method 1
of process control, but, it is directly related to process s.s. f (O O ) (2)
T k 1 k
optimization and absolute optimum point determination of output
function with respect to various input vectors. Although this where T , Ok and O k 1 are the time between two samples recorded,
approach limits the application of this control method in physical condition factors in recorded pulses k and k  1 respectively.
processes, but, due to the definition of optimum point as the best Sampling is done via a digital oscilloscope with sampling capability to
operational point of system in most processes, this method is 20 MHz as voltage respect to time diagram and the results are analyzed
widely used in control applications. When the artificial neural as a matrix form in phase systems, box tool of MATLAB software. The
networks are used as model process in control, it is not needed a reason of such high frequency-time selection respect to long pulse times
mathematical model replacement. Figure 3 shows the block is to increase machining data and guarantee the recording pulse
diagram of this control method via neural networks. capability as much as possible.

164 Archives of Computational Materials Science and Surface Engineering

EDM process optimisation via predicting a controller model

Table 1.
ED Machining parameters of sampled pulse train
Tool Copper with 10 mm dia.
Workpiece WC-Co
Machine CHARMILLES-20 ZNC
Dielectric Kerosene
Open Circuit Voltage(V) 225
Max. Allowed Current (A) 1.5,3.5,7.5,11
Pulse On - Time (µS) 35~1000
Pulse Off - Time (µS) 35~1000
Gain of Gap Magnitude(Kȍ) 120~980
Gain of Head Reaction Speed(Kȍ) 1.32~91.2
Oscilloscope Setting 100 V/div & 100 µS/div
Oscilloscope Sampling Rate 312 Samples/sec
Pulse Train Condition Factor 0.6247

a) b)

c)

Fig. 4. Membership function of: (a) discharge average voltage, (b) breakdown delay time ratio and (c) pulse condition factor as an input of
fuzzy system

The system stability factor is defined as the instantaneous machining phase analysis contains instantaneous system condition
derivative of EDM condition-time diagram if the shortest time which is a real number between zero and one, is used to create
needed to return the system to machining starting condition with couple training data of neural network. Input data are: discharge
new set up when the discharge affections with various primary voltage, discharge current (numbers of activated transistors), gap
conditions, is obliterated to the extent possibility. The results of size, head reaction speed and pulse on-off times. The neural

Volume 1 • Issue 3 • 2009 165

 R.A. Mahdavinejad

network training is based on error back propagation and via consideration. Primary data test which are later training data of
Levenberg-Marquardt. Stopping condition is defined when the neural network, also should cover space of the input variations
error is equal to 1×10-5 in training cycles. The phase sensor condition. The last obtained data as input data which leads the
elements such as preliminaries and their membership functions system to the maximum number of condition and real condition
are shown in Figure 4. Figure 5 shows the converging method of number after system parameters variation are used for training of
neural network training process. neural network and increase its knowledge. The continuous
training process of network can be stopped after certainty of any
following conditions has been established:
x The number that indicates the final condition of system is
reached to the necessity state and approved distance of desired
number.
x The network's training number and its re-optimization does
not exceed to the operator's desired number.
x For several serial optimizations, system parameters should
cause a decrease in condition number. This subject can be as a
warning for over-training of network.
Figure 7 shows the flow chart of optimization process via
predicting control based on neural network model of EDM process.

Fig. 5. Convergence diagram of supposed neural network

These factors are used to determine the system condition via

time-related functions mentioned in section 3 with the same
effecting gravity. The surface transmitted from two inputs and
single output parameters is shown in Figure 6.

Fig. 6. Fuzzy assigning surface of two input and one output

parameter

6. Optimization
6. 
Optimization
After neural network training with primary input data which is
storaged in computer's lateral memory and its corresponding
output data which is output data is done via optimization
algorithm. In complex and non-linear physical processes
optimization, since the trapping of optimization algorithm in
relativistic extremums is sometimes possible, necessity
preparations to increase the mutation and instantaneous variation
of generation conditions via generic algorithm should be under Fig. 7. Execution Flowchart of Neuro-Model Predictive Control

166 Archives of Computational Materials Science and Surface Engineering

7. Discussion
7.  and results
Discussion and Results stock removal rate in comparison with EDM system which is
set up by expert user.
x A fuzzy system is used to distinguish the condition factor of single
Two grades of WC-Co are tested in this research. Testing
pulse. In the correct parameters set up of membership functions, the
conditions are shown in Table 1. The results obtained from
input parameters have the main role in the correctness of the value
predictive controller model based on neural network and also by
obtained as the single pulse condition parameter.
expert user are shown in Figure 8.
x The input parameters such as pulse and pause times, discharge
current, gap size and reaction speed of head are analyzed and varied
0.9 dynamically during the control of system via optimizer program
0.8 Set by Controller based on genetic algorithm. In controlling structure of the system
Pulse Train Condition Factor

0.7
the escape from instability condition is under consideration.
0.6
x By hardware dismantling of controller it is possible to decrease
0.5 Set by Expert User
the cost of PC usage in control of the system in addition to
0.4
0.3
increase of the operation speed of the system. This subject causes
0.2
necessity of the decreasing in the flexibility of the system.
0.1
0
0 100 200
Time (minutes)
300 400 500 References
References
[1] A. Buchacz, Investigation of piezoelectric in influence on
Fig. 8. Condition - Time diagram of ED machining of WC-Co
characteristics of mechatronic system, Journal of Achievements
in Materials and Manufacturing Engineering 26/1 (2008) 41-48.
This Figure shows that how the carbon bridge between two
[2] A. Buchacz, Dynamical flexibility of torsionally vibrating
electrodes is distinguishable via the extremely dropped of
mechatronic system, Journal of Achievements in Materials
instantaneous condition/time curve. The stock removal rate for
and Manufacturing Engineering 26/1 (2008) 33-40.
both samples is shown in Figure 9.
[3] N. Constantin, Adaptive Neural Predictive Techniques for
0.6
Nonlinear Control, Studies in Informatics and Control 12
Set by Controller
(2003) 285-291.
Stock Removal Rate (mm^3/min)

0.5 [4] N. Constantin, I. Dumitrache, Advanced Adaptive Techniques

0.4 for Multivariable Nonlinear Processes, Proceedings of the 14th
Set by Expert User
IFAC World Congress, 1 (1999) 174-180.
0.3
[5] J.S. Donat, N. Bhat, T.J. McAvoy, Neural Net Based Model
0.2 Predictive Control, International Journal of Control 54
(1991) 1453-1468.
0.1
[6] T. Dzitkowski, A. Dymarek, Design and examining
0 sensitivity of machine driving systems with required
0 100 200 300 400 500 frequency spectrum, Journal of Achievements in Materials
Time (minutes)
and Manufacturing Engineering 26/1 (2008) 49-56.
[7] J.Y. Kao, Y.S. Tarng, A Neural Network Approach for the
Fig. 9. Stock Removal Rate - Time diagram of ED machining of WC-Co On-Line Monitoring of the Electrical Discharge Machining
Processes, Journal of Material Processing Technology 69
The structure of predictive controller model is done via PC in (1997) 112-119.
MATLAB software area with hard wary parallel contour and the [8] J. Kopac, High precision machining on high speed
desired sets up are carried out by suitable connector circuits on machines, Journal of Achievements in Materials and
EDM system. Manufacturing Engineering 24/1 (2007) 405-412.
The testing results from ED machining of WC-Co confirms [9] D. Reclik, G. Kost, J. ĝwider, The signal connections in robot
the capability of the system of predictive controller model based integrated manufacturing systems, Journal of Achievements in
on neural network with 32.8% efficiency increasing in stock Materials and Manufacturing Engineering 26/1 (2008) 89-96.
removal rate. The comparison between the set up obtained via [10] T.A. Spedding, Z.Q. Wang, Study on Modeling of Wire-
machine and expert user on EDM is done with 8 hours machining EDM Process, Journal of Material Processing Technology
that has proved the capability of the system. To improve ED 69 (1997) 18-28.
machining parameters on carbon-based materials the predictive [11] Y.S. Tarng, S.C. Ma, L.K. Chung, Determination of Optimal
controller system based on neural network is designed and Cutting Parameters in Wire- Eletrical Discharge Machining,
dismantled and the results from testing carried out on WC-Co via International Journal of Machine Tools and Manufacture 35
electro discharge sinking machine is analyzed. The following (1995) 1435-1443.
results from comparison between experimental data analysis and [12] Y.S. Tarng, S.C. Ma, P.J. Wang, K.M. Tsai, Semi-Empirical
machining data with EDM sets up via expert user are obtained: model of Work Removal and Tool Wear in Electrical
x ED sinking machine equipped with predictive controller Discharge machining, Journal of Material Processing
model of neural process is caused 32.8% improvement in Technology 114 (2001) 89-96.

READING DIRECT: www.archicmsse.org 167