Precision Engineering 47 (2017) 182–190

Contents lists available at ScienceDirect

Precision Engineering
journal homepage: www.elsevier.com/locate/precision

Improving electrical discharging machining efficiency by using

a Kalman filter for estimating gap voltages
Xue-Cheng Xi ∗ , Mo Chen, Wan-Sheng Zhao
State Key Laboratory of Mechanical System and Vibration, School of Mechanical Engineering, Shanghai Jiao Tong University, Shanghai 200240, China

a r t i c l e i n f o a b s t r a c t

Article history: Gap voltage can be used as an indicator on the direction of the electrode movement along a desired tool
Received 10 March 2016 path in electrical discharging machining (EDM) processes. However, due to the noise induced by electrical
Received in revised form 6 July 2016 discharges, the estimation of gap voltages is difficult due to the lack of an appropriate state space model.
Accepted 23 August 2016
In this paper, gap voltage signals are considered to be generated as a summation of colored noise through
Available online 28 August 2016
a linear filter and measurement noise. Obtained by the Yule–Walker auto-covariance method, the transfer
function of the linear filter can be converted into a state space model. The composite process noise and
Keywords:
the composite measurement noise are defined to derive the composite noise covariance matrices. A
Kalman filter
Stochastic process
Kalman filter can thus be designed based on the state space model and the noise covariance matrices.
Gap voltage Experimental results showed that, as compared with the traditional 10-point moving average filter, the
Electrical discharging machining (EDM) Kalman filter can decrease the average machining time as well as improve the discharging gap status.
© 2016 Elsevier Inc. All rights reserved.

1. Introduction
Electrical discharging machining (EDM) is a process which can
remove workpiece materials by a series of electrical discharges
occurring between an electrode and a workpiece [1,2]. Unlike in a
milling machining, the movement of servo axes of an EDM machine
are not only dependent on movement commands, but also deter-
mined by the gap status between an electrode and a workpiece
[3]. From the control point of view, there are two control loops in
an EDM process: the inner loop is the servo control loop, whose
function is to position the servo axes; and the outer loop is the gap
width control, whose function is to maintain an appropriate gap
width. To maintain a stable machining, the gap width between the
tool electrode and the workpiece should be kept to a distance of
10–100 ␮m [3]. As a direct measurement of the gap width during
discharging is difficult, average gap voltage [4] and ignition delay
[5] and time ratio of each gap state [6–8] are the three types of vari-
ables that can be used as feedback signals to control the gap width.
Average gap voltage is more commonly used for its simplicity in
implementation.
A gap width control block diagram when gap voltage is used to
maintain the gap width is shown in Fig. 1. Vs is the preset servo volt-
age and Vest is the filtered gap voltage. Given the voltage difference
V = Vest − Vs , a function Ke (V) is used to compute the feedrate
∗ Corresponding author.
E-mail address: xixuecheng@gmail.com (X.-C. Xi).
fr for the next servo period. Thus fr Ts is the displacement calcu-
lated by an interpolator that should be traveled during the next
servo period, where Ts is the servo period. When the gap voltage is
higher than the servo voltage (Vest > Vs ), the electrode should move
forward along the trajectory to reduce the gap width. By contrast,
when the gap voltage is lower than the servo voltage (Vest < Vs ), the
electrode should retract to increase the gap width so as to leave a
deteriorated gap condition. The integrator 1/s in the block diagram
means that the integral of the feedrate within the servo period is
the position command to the servo loop, and the gain of −1 reflects
the fact that a positive error will reduce the gap width, and vice
verse. In this paper, the existing gap width control strategy was
used, which, as shown in Fig. 2, is a nonlinear function fr = Ke (V)
relating feedrate (fr ) to gap voltage difference (V). This function
was obtained empirically by the EDM machine manufacturer.
In an ideal scenario, when a short pulse or an arc is detected, the
electrode will retract quickly so as to move away from the occur-
rence of arcs. In practice, however, it is not easy to achieve such
an ideal scenario. To keep a stable gap width, there are basically
two tasks involved: one is the design of a gap width controller, and
the other is the estimation of feedback signals, such as gap voltage
signals.
To design a gap width controller, it is important to build a
dynamic model for EDM processes. Due to the difficulties in mod-
eling EDM processes, fuzzy logic [9–11], neural networks [12,13],
and genetic algorithm [14] are often used to model the dynam-
ics of EDM processes and to help to design a gap width controller.
Many researchers developed various servo controllers for EDM

http://dx.doi.org/10.1016/j.precisioneng.2016.08.003
0141-6359/© 2016 Elsevier Inc. All rights reserved.
 X.-C. Xi et al. / Precision Engineering 47 (2017) 182–190 183

transfer function, or an autoregressive model, is built by using the

Fig. 1. Block diagram of gap width control for EDM processes.
Yule–Walker auto-covariance method. A discrete–time stochastic
state space model is then obtained from the transfer function. To
alleviate the online computational burden, a steady-state Kalman
filter is used. A steady-state gain from the measurement to the esti-
mated signal is calculated so that the estimated value can be scaled
back to the original range of the measurements. By collecting gap
voltage signals from an EDM test, a Kalman filter can be built for
the EDM machine. The filtered gap voltage by the Kalman filter is
then used as the feedback signal to the gap width controller. Exper-
imental tests showed that, the machining time by using the Kalman
filter is less than that by using the traditional moving average filter.
The Kalman filter is also found to be able to improve the gap status.

2. Modeling of gap voltage dynamics

Kalman filter is an optimal linear filter which estimates a sig-

nal from a noisy environment in a minimal statistical error sense.
Considering the fact that noise contained in measured gap voltage
signals, it is helpful to use a Kalman filter to optimally estimate the
gap voltage. To design a Kalman filter, a state space model is needed.
To this end, the gap voltage signal y(k) is regarded as a summation
of two components

y(k) = y0 (k) + v(k) = Gw (q−1 )w(k) + v(k) (1)

where w(k) is the process noise, v(k) is the measurement noise,

Gw (q−1 ) is a linear filter, and q−1 is the unit delay operator. It is
Fig. 2. Nonlinear function relating feedrate to gap voltage difference for gap width
first assumed that both w(k) and v(k) are white noise, and are
control.
uncorrelated with each other, i.e., E[w(k)v(k)] = 0, and w(k) and
v(k) have the same variance, i.e., w2 = v2 . It is noted from Eq. (1)
processes. Due to its time-varying and stochastic nature, an online that y0 (k) is the colored noise obtained by feeding Gw (q−1 ) with
identification algorithm was used to obtain the process model, and the process noise w(k). To obtain the transfer function Gw (q−1 ), the
then a controller was designed based on the model parameters Yule–Walker method can be used [22] by using the signal sequence
[6,7,15,16]. Wu et al. used an adaptive control system to directly y0 (k).
control electrode discharging cycles [17]. A variable structure sys- The transfer function Gw (q−1 ) from w(k) to y0 (k) can be written
tem (VSS) controller was reported to control the gap width and as
achieve a good machining performance [18] and later robust gap
y0 (k) b0
width control was applied to compensate for the electrode wear in Gw (q−1 ) = =
an electric discharge scanning (ED-Scanning) process [19]. w(k) 1 + a1 q−1 + a2 q−2 + · · · + an−1 q−(n−1) + an q−n
The second task in the design of a gap controller is the esti- (2)
mation of feedback signals. In this paper gap voltage is chosen
where n is the order of the transfer function, ai , i = 1, . . ., n are the
as the feedback signal. When making a judgment on whether to
coefficients of the denominator, and b0 is a constant which can be
move forward or backward, the gap voltage is a crucial measure.
chosen such that the steady-state gain is unity.
As EDM machining is a stochastic process, there exists strong noise
To the use the Yule–Walker method, it is necessary to use the
in measured gap voltage signals. In the gap width control loop as
covariance matrix R0 of y0 (k), which is
shown in Fig. 1, a filter or an estimator is designed to estimate the
gap voltage from a noisy environment. To minimize the estima- ⎡ ⎤
r0 (0) r0 (1) ··· r0 (M − 1)
tion error in a statistical sense, an optimal filter can be used to
⎢ r0 (1) r0 (0) ··· r0 (M − 2) ⎥
⎥
remove the noise components. Conventional EDM machines sim- ⎢
ply use moving average filters. Though working smoothly, a moving R0 = ⎢
⎢
⎥
⎥ (3)
.. .. .. ..
average filter cannot be considered as an optimal estimator for gap ⎣ . . . . ⎦
voltages. r0 (M − 1) r0 (M − 2) ··· r0 (0)
Kalman filters have long been considered as optimal estimators
for noisy signals [20–22]. Kalman filters are designed based on state where r0 (k) = E[y0 (t)y0 (t + k)], and E is the expectation operator.
space models. Though a Kalman filter can be applied in a stochastic However, the colored noise y0 (k) is unavailable. What is available
EDM machining process, the building of a state space model for a is the measured gap voltage y(k), which contains both the colored
Kalman filter is not a trivial task. Give a sequence of noisy signals, noise y0 (k) and the measurement noise v(k). The covariance r(k) of
the challenge for the design of a Kalman filter is to build a state space y(k) can be computed as
model. Zhou et al. [23] used two interactive Kalman filters based
on an instrumental variable approach to obtain unbiased model r(k) = E[y(t + k)y(t)] = E[(y0 (t + k) + v(t + k))(y0 (t) + v(t))]
parameter estimates online for an EDM process.
= E[y0 (t + k)y0 (t) + y0 (t + k)v(t) + v(t + k)y0 (t) + v(t + k)v(t)]
In this paper, a Kalman filter is designed and used together with

a gap width controller to form a closed loop as illustrated in Fig. 1. r0 (0) + v2 , k=0
A state space model for a Kalman filter is first built based on an = (4)
autoregressive (AR) model for gap voltage signals. A discrete–time r0 (k), k=
/ 0
184 X.-C. Xi et al. / Precision Engineering 47 (2017) 182–190

∼ r0 (0), mean-
In this paper, it is assumed that r(0) = r0 (0) + v2 = where x(k) is the state vector. The matrices of A, C, G and H are
ing that the measurement noise is negligible as compared with the obtained from the transfer function Gw (q−1 ) from w(k) to y0 (k) as
gap voltage. With this assumption and Eq. (4), the covariance matrix in Eq. (2). A, C, G and H are written, respectively, as
of R0 can be approximated by R, i.e., ⎡ ⎤
−a1 −a2 ··· −an−1 −an
⎡ ⎤ ⎢ ⎥
r(0) r(1) ··· r(M − 1) ⎢ 1 0 ··· 0 0 ⎥
⎢ r(M − 2) ⎥
⎢ ⎥
⎢ r(1) r(0) ···
⎥ A=⎢
⎢
0 1 ··· 0 0 ⎥
⎥ (13)
R0 ∼
=R=⎢ .. .. .. .. ⎥ (5) ⎢ ⎥
⎣ ⎦ .. .. .. .. ..
. . . . ⎣ . . . . . ⎦
r(M − 1) r(M − 2) ··· r(0) 0 0 0 1 0
⎡ ⎤
To compute the coefficients of the transfer function Gw (q−1 ), the 1
Yule–Walker equation is used as ⎢0⎥
⎢ ⎥
G=⎢ . ⎥ (14)
⎡ ⎤ ⎡ c1 ⎤ ⎡ ⎤ ⎣ .. ⎦
r(0) r(1) · · · r(M − 1) r(1)
⎢ r(1) r(0)
⎢
· · · r(M − 2) ⎥ ⎢ c2
⎥ ⎢ r(2) ⎥ 0
⎢ ⎥⎢ ⎥ ⎢
⎢ ⎥⎢ ⎥=⎢ . ⎥

⎣
.. .. .. ..
⎦ ⎣ .. ⎥ ⎣ . ⎥ (6)
C = −b0 a1 −b0 a2 ··· −b0 an−1 −b0 an
. ⎦
(15)
. . . .
. ⎦
r(M − 1) r(M − 2) ··· r(0) r(M) H = b0 (16)
cM
It is noted that the conversion from a transfer function to a space
where ci are the coefficients that need to be solved. Eq. (6) can also model is not unique. Here in this paper, the canonical form as listed
be written in a compact matrix form as in Eqs. (13)–(16) is adopted (P85 of [21]). H =/ 0 because the orders
of the numerator and denominator of Gw (q−1 ) in Eq. (2) are both
Rc = r (7) equal to n in terms of q.
For the process noise w(k) and the measurement noise v(k), the
where R is the covariance matrix as in Eq. (5) and the coefficient two covariance matrices, Q(k) and R(k), and the cross-covariance
vector c can be obtained by matrix, N(k), are defined as

c = R−1 r (8) Q (k) = E[w(k)wT (k)] (17)

T
R(k) = E[v(k)v (k)] (18)
where
N(k) = E[w(k)vT (k)] = 0 (19)
c = [c1 , c2 , . . ., cM ] = [−a1 , −a2 , . . ., −aM ] (9)
One note about Eq. (19) is that, as the process noise w(k) and the
After obtaining the gap voltage dynamic model in terms of a measurement noise v(k) are uncorrelated, therefore N(k) = 0.
transfer function as in Eq. (2), it is beneficial to check the whiteness To facilitate the deviation of the Kalman filter, the composite
of the process noise w(k) to Gw (q−1 ). By neglecting the measure- process noise w̄(k) and the composite measurement noise v̄(k) are
ment noise v(k), w(k) can be approximated by feeding the inverse defined, respectively, as
−1 −1
transfer function Gw (q ) with the gap voltage signal as illustrated
w̄(k) = Gw(k) (20)
by
v̄(k) = Hw(k) + v(k) (21)
1
w(k) ∼
= (1 + a1 q−1 + a2 q−2 + · · · + an−1 q−(n−1) + an q−n )y(k)
b0 With the definitions of w̄(k) and v̄(k), Eq. (12) can be written as
(10)
x(k + 1) = Ax(k) + w̄(k)
(22)
y(k) = Cx(k) + v̄(k)
White noise is characterized by an autocorrelation function as
The composite process noise covariance matrix Q̄ (k) is derived
defined by
as

2,
w k=0 Q̄ (k) = E[Gw(k)wT (k)GT ] = GQ (k)GT (23)
E[w(t)w(t + k)] = (11)
0, k=
/ 0 The composite measurement noise covariance matrix R̄(k) is
derived as

3. Design of a Kalman filter for EDM processes R̄(k) = E[(Hw(k) + v(k))(wT (k)H T + vT (k))]

= E[Hw(k)wT (k)H T + Hw(k)vT (k) + v(k)wT (k)H T + v(k)vT (k)]

3.1. Conversion from transfer function to state space model
= HQ (k)H T + R(k) (24)
For the design of a Kalman filter, the state space model can be
written as The composite cross-covariance matrix N̄(k) between w̄(k) and v̄(k)
is obtained as
x(k + 1) = Ax(k) + Gw(k) T T
N̄(k) = E[Gw(k)(Hw(k) + v(k)) ] = E[Gw(k)w(k) H T + Gw(k)vT (k)]
y(k) = Cx(k) + Hw(k) + v(k) (12)
= y0 (k) + v(k) = GQ (k)H T (25)
 X.-C. Xi et al. / Precision Engineering 47 (2017) 182–190
With the state space model Eq. (12) and the three composite
noise covariance matrices(Eqs. (23)–(25)), a Kalman filter can be
designed based these information.
At the current instant k, the calculations of a Kalman filter can
be separated into two stages: the measurement update stage and
the time update stage [20]. During the measurement update stage,
a newly obtained measurement y(k) at instant k is used to update
the a priori state vector x̂(k|k − 1) which was computed at instant
(k − 1), thus the a posteriori state vector x̂(k|k) is computed as
x̂(k|k) = x̂(k|k − 1) + M(k)(y(k) − C x̂(k|k − 1)
v̂(k|k) = Kv (k)(y(k) − C x̂(k|k − 1))
x̂(k + 1|k) = Ax̂(k|k) + v̂(k|k) = Ax̂(k|k − 1) + L(k)(y(k) − C x̂(k|k − 1))
where the gain matrices M(k), Kv (k), and L(k) are defined as
−1
M(k) = P(k|k − 1)C T (CP(k|k − 1)C T + R̄(k))
−1
Kv (k) = N̄(k)(CP(k|k − 1)C T + R̄(k))
T
L(k) = AM(k) + Kv (k) = (AP(k|k − 1)C
−1
+ N̄(k))(CP(k|k − 1)C T + R̄(k))
The calculations of M(k), Kv (k) and L(k) need the error covariance
matrix P(k|k − 1), whose update is governed by the Riccati equation
as
T
P(k + 1|k) = AP(k|k − 1)A + Q̄ (k)
 T
185
3.3. Steady-state gain from y(k) to ŷ(k)
The raw measured data of the gap voltage from the A/D device of
the EDM machine is in integer forms. To avoid possible numerical
instability problems in matrix calculations, the raw voltage data
is preprocessed by removing the DC component and then scaled
down to a range of [−0.5, 0.5]. However, it was found that the output
range of the steady-state Kalman filter is not in the same range
as the input. As the gap voltage is used for judging the electrode
movement directions, it is desirable that the range of the output
(26) of the steady-state Kalman filter is the same as that of the input.
(27) To this end, the steady-state Kalman filter can be considered as a
linear system, in which the measured signal y(k) is the input, and
the filtered output ŷ(k) and the estimated state vector x̂(k) are the
outputs. The task here is to derive the transfer function Gyŷ (z) from
(28) y(k) to ŷ(k). The state space model in steady state is written as
x̂(k|k) = x̂(k|k − 1) + M(y(k) − C x̂(k|k − 1) (35)
x̂(k + 1|k) = Ax̂(k|k − 1) + L(y(k) − C x̂(k|k − 1) (36)
(29)
ŷ(k|k) = C x̂(k|k)) (37)
(30)
First, taking the z-transform to Eqs. (35) and (37) gives
X̂k|k (z) = (I − MC)X̂k|k−1 (z) + M Ŷk (z) (38)
(31) Ŷk|k (z) = C X̂k|k (z) = C(I − MC)X̂k|k−1 (z) + CMY k (z) (39)
Taking the z-transform to Eq. (36) gives
X̂k+1|k (z) = AX̂k|k−1 (z) + L(Yk (z) − C X̂k|k−1 (z)) (40)
X̂k|k−1 (z) can then be expressed by
X̂k|k−1 (z) = (zI − A + LC)−1 LY k (z) (41)
 T
− L(k) CP(k|k − 1)C + R̄(k) L (k) (32) Substituting Eq. (41) into Eq. (39) gives

Yˆk (z)
To start with the recursive calculation, the initial condition of = C(I − MC)(zI − A + LC)−1 L + CM (42)
Yk (z)
P(k|k − 1) is used as
−1
= C̄(zI − Ā) B̄ + D̄ (43)
P(1|0) = pI (33)
Ny (z)
where I is an identity matrix of n-order, p is a scalar. Here p is chosen = (44)
Dy (z)
as 100 in this paper.
where Ā = A − LC, B̄ = L, C̄ = C(I − MC), and D̄ = CM. Ny (z) and
3.2. Steady-state Kalman filter Dy (z) are the numerator and denominator of the transfer function
Yˆk (z)/Yk (z), respectively. By setting z = 1, the steady-state gain from
y(k) to ŷ(k) can be obtained as

P = APAT + Q̄ − L CPC T + R̄ LT
Ny (1)
Dy (1)
−1
= APAT + Q̄ − (APC T + N̄)(CPC T + R̄) (CPAT + N̄ T )
For an EDM machining process which usually lasts for a long
time, the error covariance matrix P(k) will reach a constant matrix
in a steady state after a transient period. To alleviate the compu-
tational load in a real-time system, a steady-state Kalman filter
can be actually implemented, where the constant gain matrices L
and M are used. While the loss resulting from the use of a steady-
state Kalman filter is negligible, the gain by using a steady-state
Kalman filter is significant for a real-time control system as matrix
calculations have been largely reduced, especially for high order
systems.
The steady-state Kalman filter can be derived from the time-
varying ones, with the time-varying error covariance matrix P(k)
replaced by a constant matrix P as in Eq. (34). The Matlab function
kalman can be used to compute the matrics L, M and P in steady
state.
Yˆk (z)
Gyŷ = = (45)
Yk (z)
z=1
(34)
For easy reference, the flow chart of the implementation of a
steady-state Kalman filter is illustrate in Fig. 3, where, yread (k) and
yest (k) are the reading and estimate of the gap voltage, respectively,
scale() is a function for scaling a voltage reading to a value in [−0.5,
0.5] and scaleback() is a function for scaling a value in [−0.5, 0.5]
back to the range of voltage readings.
4. Experimental results
4.1. Modelling of gap voltage dynamics
A series of machining tests were carried out on an HE70 EDM
machine (Shanghai HanSpark Co.). The schematic of the exper-
imental setup is shown in Fig. 4. For this study, the computer
numerical control (CNC) system of the EDM machine was devel-
oped by Shanghai Jiao Tong University and the original CNC system
was bypassed. The EDM machine is controlled by an industrial PC
186 X.-C. Xi et al. / Precision Engineering 47 (2017) 182–190

Fig. 5. (a) Original gap voltage reading; (b) 10 point moving average of gap voltage.

voltages and the corresponding filtered values by the moving aver-

Fig. 3. Flow chart of the implementation of a steady-state Kalman filter.
Fig. 4. Experimental EDM CNC machine.
age filter are shown in parts (a) and (b) of Fig. 5, respectively.
In order to design a Kalman filter, a series of gap voltages were
collected in an EDM machining test as shown in Fig. 6(a). An order of
n = 10 is chosen by trial and error. By using the Yule–Walker equa-
tion (Eq. (6)), the coefficients are computed and listed in Table 1.
This model was obtained when the material of the workpiece is tool
steel and that of the electrode is Poco EDM-C3.
With the coefficients in Table 1, a Kalman filter is designed
according to the flow chart in Fig. 3. The Kalman filter is then used
to filter the gap voltages and the filtered gap voltages are shown in
Fig. 6(b). It can be seen that the noisy components in gap voltages
have been largely removed. As a comparison, the outputs from the
moving average filter as shown in Fig. 5(b) still contain noticeable
noisy components.
The time shifted covariance, r(k), is plotted in Fig. 7. It can be seen
that there is an impulse when the time difference k = 0, while the
time covariance r(k) when k = / 0 is very small as compared to the
value of r(0), which can be effectively considered as uncorrelated in
practice. Satisfying the condition as specified in Eq. (11), the input
signal w(k) to the filter of Eq. (2) can be considered as white noise.

with a PMAC PCI motion control card. The CNC software is run-
ning on Ubuntu 10.04 with the RTAI real-time patch. The real-time
sampling period on the PC is set to 2 ms and the servo period on
PMAC is 0.88 ms. The real-time communication between PC and
PMAC is achieved through the dual-ported RAM (DPRAM) on the
PMAC PCI card. The raw gap voltage is firstly measured by an ACC-
28A DAC card connected with the PMAC card. The voltage DAC
reading is then sent to the PC through DPRAM.
The original filter for gap voltages used in the EDM machine is
a 10-point moving average filter which is defined as


9
y(k) = ai u(k − i)
i=0
(46)
10 − i
ai = , i = 0, . . ., 9
10

where the choice for the coefficients ai are based on experiences.

The moving average filter is implemented by a PLC program run-
ning in PMAC with a sampling period of 0.88 ms. The original gap Fig. 6. (a) Original gap voltage; (b) Kalman filter output.
 X.-C. Xi et al. / Precision Engineering 47 (2017) 182–190 187

Table 1
y(k) b0
Dynamic model of gap voltage, w(k) = Ts =2 ms.
A(q−1 )

b0 E[w2 (k)] Gyŷ

0.291503249296 0.036941377759 2.822878154362

A(q−1 ) q0 q−1 q−2 q−3 q−4

1.000000000000 −0.197521057202 −0.095901307528 −0.086399772122 −0.060585322073
q−5 q−6 q−7 q−8 q−9 q−10
−0.062497298750 −0.047383331386 −0.011782011498 −0.082531158936 −0.070773327350 0.006877836140

Table 3
Machining test by 10-point moving average (hole depth = 5 mm, diameter = 10 mm).

No Machining time (s) Retraction time ratio Carbon deposit

1 810.388 0.1973 Yes

2 738.396 0.1858 No
3 905.180 0.2139 Yes
4 810.076 0.2047 No
5 650.824 0.1637 No

Average 782.973 0.1931 –

to Table 2, the set of machining parameters in the second column

Fig. 7. Auto correlation of inverse filtered gap voltage.
4.2. Machining tests by two filters
Moving average filters are usually adopted for its simplicity in
conventional EDM machines. The moving average filter used in this
paper is defined in Eq. (46).
To reflect the time that is used for electrode retraction, the
retraction time ratio  is defined as
Tretract
=
Ttotal
where Tretract is the time used for retracting from deteriorated gap
conditions, and Ttotal is the total machining time excluding the elec-
trode jump time.
To compare the machining performance with the use of the 10-
point moving average and the Kalman filter, a series of machining
tests were carried out to machine by EDM a hole by using an elec-
trode with a diameter of 10 mm. Two hole depths tested are 5 mm
and 10 mm, respectively. For each depth, a comparison of machin-
ing performance was made between the moving average filter and
the Kalman filter. For a fair comparison, the same set of machin-
ing parameters is used for the same hole depth as listed in Table 2.
These machining parameters include servo voltage, peak current,
pulse on time, pulse off time, jump time, and jump period. Referring
is used when the hole depth is 5 mm, and the set of machining
parameters in the third column is used when the hole depth is 10
mm.
When machining the 5mm-deep holes, the material of the elec-
trode used is Poco EDM-C3, which contains copper, and that of the
workpiece is tool steel. Referring to Fig. 1, the difference between
these two machining tests is the choice of a filter for gap voltages.
One test used the 10-point moving average filter by Eq. (46) and
the other used the Kalman filter. Machining time and retraction
time ratio are recorded. Each filter was tested for five times and the
results are shown in Tables 3 and 4.
It can be noted from Tables 3 and 4 that the average machining
time has decreased by 25.7 % from 782.973 seconds by using the
10-point moving average to 581.656 seconds by using the Kalman
filter. This can be explained by the fact that there are less fluctu-
ations in the estimated voltage signals, there is a decrease in the
time used for first moving back when the voltage is below the servo
reference voltage and then moving forward when the voltage is
beyond the servo reference voltage. The time spent on unneces-
(47)
sary back and forth movements has been reduced, and more time
has been spent on real discharges for removing materials. It can also
be noticed from Tables 3 and 4 that the averaged retraction time
ratio has been reduced from 19.6 % by the moving average filter to
16.3 % by the Kalman filter.
The decrease in retraction time gives one reason for the decrease
in the total machining time. Another reason is the improved gap
status by the use of a Kalman filter. Two holes machined by the
moving average and the Kalman filter are shown in parts (a) and
(c) of Fig. 8, respectively. It can be seen from Fig. 8(a) that on
the bottom of the hole machined by the moving average filter,
there are black carbon deposits, which result from unstable dis-
charges. By contrast, as shown in Fig. 8(c), there is no carbon deposit
on the bottom of the hole machined by the Kalman filter. Accu-
mulated carbon deposits are harmful for stable discharges. These

Table 2
EDM machining parameters.
Table 4
Hole depth(mm) 5 10 Machining test by Kalman filter (hole depth = 5 mm, diameter = 10 mm).
Electrode Poco EDM-C3 Poco EDM-4
No Machining time (s) Retraction time ratio Carbon deposit
Workpiece Tool steel Tool steel
1 636.790 0.1759 No
Open voltage (V) 150 150
2 537.622 0.1628 No
Peak current (A) 28 28
3 444.402 0.1489 No
Pulse on (␮m) 60 60
4 756.240 0.1782 No
Pulse off (␮m) 40 40
5 533.226 0.1502 No
Up time (s) 0.6 1
Down period (s) 5 5 Average 581.656 0.1632 –
188 X.-C. Xi et al. / Precision Engineering 47 (2017) 182–190

Fig. 9. Gap voltage signals at the depth of 4mm with 10-point moving average filter:
(1) measured gap voltage; (2) moving average filter.

carbon deposits on the bottom of the hole. By contrast, there is

no carbon deposit found at the bottoms of all holes machined by
the Kalman filter, as illustrated in the 4th column of Table 4. It is
thus concluded that the Kalman filter not only reduces unneces-
sary retraction movements, but also improves the discharging gap
status.
It was found that the machining speed is significantly reduced
when the hole depth is more than 3 mm. The gap voltage sig-
nals when the depth is about 4 mm were recorded for both the
moving average filter and the Kalman filter, and the results are
shown in Figs. 9 and 10, respectively. It was found that the fre-
quency of oscillations in the gap voltages by using the moving
average filter is higher than that by using the Kalman filter. With
the increase in the hole depth, the flushing conditions become
more severe than at the initial stage where the flushing condi-
tions are better. With more oscillations in gap voltages resulting
from the poor gap status, it is important to obtain a good esti-
mate of the gap voltage so as to avoid unnecessary back and forth
movements, which, in turn, could deteriorate the gap status even
further.

Fig. 8. Machined holes by: (a) moving average filter (the bottom has carbon
deposit); (b) moving average filter (the circular black area are resulted from arcs);
(c) Kalman filter.

accumulated carbon deposits can be easily attached to the elec-

trode, thus forming a carbon cone on it. As discharges are more
likely to occur at the tip of a cone, arcs have a higher probability
of occurence. As illustrated in Fig. 8(b), the circular black area on
the bottom of a hole was resulted from arcs. With the occurrence
of continuous arcs, the EDM machining has to be stopped to avoid
damages to the electrode and the workpiece.
The observation results for carbon deposits are summarized in
the 4th column of Tables 3 and 4. It can be noticed that, for the Fig. 10. Gap voltage signals at the depth of 4mm with Kalman filter: (1) measured
holes machined by the moving average fitler, 3 out of 5 holes have gap voltage; (2) Kalman filter.
 X.-C. Xi et al. / Precision Engineering 47 (2017) 182–190 189

Fig. 11. Comparison of machining time including extraction time for a hole with depth = 10 mm and diameter = 10 mm.

Table 5 5. Conclusion
Comparison of machining time (excluding electrode jumping time) for a hole with
depth = 10 mm, diameter = 10 mm).
Estimation of gap voltages is crucial for gap width control in EDM
Filter Machining Retraction time Carbon deposit processes. However, strong noise induced by electrical discharges
time (s) ratio (%)
between an electrode and a workpiece makes the estimation of
4278.314 0.2383 Yes gap voltages a difficult task. Considering the stochastic nature of an
Moving average 3613.214 0.2293 Yes EDM process, a Kalman filter is used to estimate gap voltages in an
3733.266 0.2231 Yes
optimal statistical error sense.
Average 3874.931 0.2303 – Gap voltage signals are thought of being generated as a sum-
3468.318 0.2166 No mation of colored noise through a linear filter and measurement
Kalman filter 3346.290 0.2107 No noise. The transfer function of the filter, with the same order for
3148.254 0.2038 No the numerator and denominator, is built up by the Yule–Walker
Average 3320.954 0.2103 – covariance method. A state space model is then derived from the
transfer function. The composite process noise and the compos-
ite measurement noise are used in the derivation of the composite
A hole with a depth of 10 mm was also machined to compare the
noise covariance matrices. With the state space model as well as
machining performance when the moving average and the Kalman
the covariance matrices, a Kalman filter can thus be designed. Scal-
filter were used, respectively. In this test, the material of the elec-
ing is both needed in the pre-processing and post-processing of the
trode has been changed to Poco EDM-4, which does not contain
Kalman filter. A steady-state gain from the measured gap voltage
copper and the material of the workpiece is still tool steel. Though
to the output of a Kalman filter is used in the post-processing.
the material of the electrode has been changed, the same Kalman
Experimental results show that, as compared with the 10-point
filter based on the model as listed in Table 1 is used, where the
moving average filter, the Kalman filter can reduce the average
model was obtained when the material of the electrode is Poco
machining time. When machining a 5 mm-deep hole, the reduction
EDM-C3. To cope with the increase of the hole depth, the electrode
percentage in average machining time is 25.7%. When machining a
jump speed is increased. The machining conditions for the 10-mm
10 mm-deep hole, the reduction percentage in average machining
hole are listed in the third column of Table 2. For each filter, three
time is 14.3%. The use of the Kalman filter is also found to be able
tests were carried out and the results are listed in Table 5. It can
to improve the discharging gap status.
be noted from Table 5 that the average machining time by using
The Kalman filter designed in this paper has shown its advan-
the Kalman filter is 14.3% less than that by the 10-point moving
tages over the traditional moving average filter. Future research
average filter.
will focus on adaptive Kalman filters for estimating gap voltages
Fig. 11 shows the two curves of the machining time with respect
to cope with varying machining conditions, and the design of an
to the hole depth when the moving average filter and the Kalman
advanced gap width controller.
filter are used, respectively. It can be noticed that at discrete values
of the hole depth ranging from −1 mm to −10 mm, the machining
times by using the Kalman filter are all less than those by using the Acknowledgements
moving average filter. This is because with the increase in the hole
depth, the gap conditions become deteriorated, and the machining This research is financially supported by National Natural
time for each single millimeter has been elongated. The advantage Science Foundation of China (No. 51675340), National Science
of the Kalman filter over the moving average filter becomes more and Technology Major Project(No. 2014ZX04001061), Gas Tur-
obvious with the increase in hole depth. bine Research Institute, Shanghai Jiao Tong University (No.
190 X.-C. Xi et al. / Precision Engineering 47 (2017) 182–190
AF0200088/015), and Industry-Academia-Research Project of
Shanghai Municipal Education Commission (No. 15CXY03).
Appendix A. Supplementary Data
Supplementary data associated with this article can be found,
in the online version, at http://dx.doi.org/10.1016/j.precisioneng.
2016.08.003.
References
[1] Ho K, Newman S. State of the art electrical discharge machining (EDM). Int J
Mach Tools Manuf 2003;43(13):1287–300.
[2] Abbas NM, Solomon DG, Bahari MF. A review on current research trends in
electrical discharge machining (EDM). Int J Mach Tools Manuf
2007;47(7):1214–28.
[3] Kunieda M, Lauwers B, Rajurkar K, Schumacher B. Advancing EDM through
fundamental insight into the process. CIRP Ann-Manuf Technol
2005;54(2):64–87.
[4] Fujiki M, Kim G-Y, Ni J, Shih AJ. Gap control for near-dry EDM milling with
lead angle. Int J Mach Tools Manuf 2011;51(1):77–83.
[5] Behrens A, Ginzel J. A comparison of different input values for gap-width
controllers used in electro-discharge machining. In: Proceedings of 13th
international symposium for electro machining (ISEMXIII). 2001. p. 93–107.
[6] Rajurkar K, Wang W, Lindsay R. A new model reference adaptive control of
EDM. CIRP Ann-Manuf Technol 1989;38(1):183–6.
[7] Rajurkar K, Wang W, Lindsay R. Real-time stochastic model and control of
EDM. CIRP Ann-Manuf Technol 1990;39(1):187–90.
[8] Rajurkar K, Wang W. Improvement of EDM performance with advanced
monitoring and control systems. J Manuf Sci Eng 1997;119(4B):770–5.
[9] Boccadoro M, Dauw D. About the application of fuzzy controllers in
high-performance die-sinking EDM machines. CIRP Ann-Manuf Technol
1995;44(1):147–50.
[10] Behrens A, Ginzel J. Neuro-fuzzy process control system for sinking EDM. J
Manuf Process 2003;5(1):33–9.
[11] Kao C-C, Shih AJ, Miller SF. Fuzzy logic control of microhole electrical
discharge machining. J Manuf Sci Eng 2008;130(6):064502.
[12] Kao J, Tarng Y. A neutral-network approach for the on-line monitoring of the
electrical discharge machining process. J Mater Process Technol
1997;69(1):112–9.
[13] Tsai K-M, Wang P-J. Predictions on surface finish in electrical discharge
machining based upon neural network models. Int J Mach Tools Manuf
2001;41(10):1385–403.
[14] Mandal D, Pal SK, Saha P. Modeling of electrical discharge machining process
using back propagation neural network and multi-objective optimization
using non-dominating sorting genetic algorithm-II. J Mater Process Technol
2007;186(1):154–62.
[15] Wang WM, Rajurkar KP, Akamatsu K. Digital gap monitor and adaptive
integral control for auto-jumping in EDM. J Eng Ind 1995;117(2):253–8.
[16] Zhou M, Han F. Adaptive control for EDM process with a self-tuning regulator.
Int J Mach Tools Manuf 2009;49(6):462–9.
[17] Wu J, Zhou M, Xu X, Yang J, Zeng X, Xu D. Fast and stable electrical discharge
machining (EDM). Mech Syst Signal Process 2016,
http://dx.doi.org/10.1016/j.ymssp.2015.11.006.
[18] Chang Y-F. VSS controller design for gap control of EDM. JSME Int J Ser C Mech
Syst Mach Elem Manuf 2002;45(3):712–21.
[19] Chang Y-F, Chiu Z-H. Electrode wear-compensation of electric discharge
scanning process using a robust gap-control. Mechatronics
2004;14(10):1121–39.
[20] Åström KJ, Wittenmark B. Computer-controlled systems: theory and design.
3rd ed. Upper Saddle River, NJ: Prentice Hall; 1997.
[21] Franklin G, Powell J, Workman M. Digital control of dynamic systems.
Addison-Wesley; 1998.
[22] Haykin S. Adaptive filter theory. 4th ed. Upper Saddle River, NJ: Prentice Hall;
2002.
[23] Zhou M, Meng X, Qin J, Chen Z, Lian X. Building an EDM process model by an
instrumental variable approach based on two interactive Kalman filters.
Precis Eng 2013;37(1):146–58.