Small-Signal Model of a 5kW High-Output Voltage Capacitive-Loaded Series-Parallel

Resonant DC-DC Converter

Fabiana da Silveira Cavalcante and Johann W. Kolar
Swiss Federal Institute of Technology (ETH) Zurich
Power Electronic Systems Laboratory
ETH Zentrum / ETL I14, Physikstrasse 3
CH-8092 Zurich / SWITZERLAND
[email protected]
[email protected]

Abstract This paper presents an accurate method, based on first

harmonic analysis, to derive the small-signal model of the series-parallel
resonant DC-DC converter with capacitive output filter for high-voltage
applications. The model provides a fast and reliable way to obtain the
desired transfer functions of the system for each operating point. For this
reason, the small-signal model simplifies the controller design task for
resonant converters by removing the need of complex mathematical
analysis and saves the time required for measurements when using trial
and error design method. The small-signal model has been verified
through simulation and experimental measurements.

I. INTRODUCTION
Series-parallel resonant converters are frequently employed for
the realization of high output voltage DC-DC converters since they
can integrate the non-idealities of the high-voltage transformer.
However, the design of resonant converters is involved due to the
large number of operating states occurring within a pulse period
[1]. Specifically, the controller design task is involved because the
small-signal transfer function is extremely dependent on the
operating point of the converter. There are a large number of
representative publications on modeling and control of resonant
converters where small-signal models for the series, parallel and
inductive loaded series-parallel resonant converters are proposed
[2,4,5,6]. However the small-signal model of the capacitive loaded
series-parallel resonant converter is still missing from the literature
and for this reason many controller designs are performed by trial
and error [7]. The trial and error procedure is not convenient
because the small-signal model varies significantly with load
changes and the load is normally an unknown parameter.
Therefore, the setting of the controller parameters can take a long
time until a robust set of parameters are found.

This paper develops a small-signal model for the series-parallel

resonant DC-DC converter with capacitive output filter in order to
make the controller design for this type of converter easier. The
converter power circuit is initially described and an analytical
description of the transient and steady-state behavior is given
based on the generalized averaging method [2]. Furthermore, the
small-signal model of the converter is determined and the
magnitude and phase plots of the transfer function of the control
variable to the output voltage are shown. Results from simulations
of a 5kW converter with output voltage in the range from 23kV to
62.5kV and operating at 250 kHz at full load and 500 kHz at low
load, are given. The simulation results fully verify the theoretical
considerations. Finally, the accuracy of the small-signal model is
verified by a laboratory prototype.
II. CIRCUIT DESCRIPTION
The topology of the 5kW series-parallel resonant full-bridge
DC-DC converter with impressed output voltage, operating above
resonance, is shown in Fig. 1. The (parasitic) capacitors C1, C3
serve for zero-voltage switching of power transistors S1 and S3, the
resonant inductor Ls is formed by the transformer stray inductance
in combination with an auxiliary inductor connected in series. The
parallel resonant capacitor Cp is formed by the parasitic
capacitance of the high-voltage transformer T1 secondary winding.
The converter output power is controlled by varying the duty-cycle
and the operating frequency is automatically adjusted to ensure the
commutation of one bridge leg occurs at zero current. This bridge
leg is realized with IGBTs and additional diodes, D2, and D4. The
second bridge leg commutates at zero voltage, therefore power
MOSFETs are employed and the free-wheeling diodes D1, D3 are
the intrinsic diodes of the MOSFETs.
T1

S1

D1

C1

S2

D2

Vin

i Ls
B

Do1

Io
Ch1

Cs

Ls
Do2

Cp

Ch2
Ro

S3

D3

C3

S4

D4

Do3

Vo

Ch3

1:n
Do4

Ch4

Fig. 1: Structure of the power circuit of a series-parallel resonant DC/DC converter with impressed output voltage; Cp denotes the equivalent capacitance of the
secondary winding referred to the primary.

S1

D1

S2

D2

Vin

Cs

I Ls

D3

S4

Io

IT
Cp

VCs

B
S3

Ls

Do1

VCp

Co1'

Ro V
o

D4

Co2'

Do2

Fig. 2: Simplified circuit of the converter with the output components referred to the transformer primary side.

In order to simplify the analysis it is assumed that: all the

components are ideal and have no losses; the voltage, Vin, supplied
by the input pre-regulator is constant and has no ripple; the
magnetizing inductance of the transformer is not considered in the
analysis; capacitors C1,C3, are too small so they are not included in
the analysis; the load is modeled as a pure resistance and all the
output quantities are referred to the transformer primary side. From
these assumptions the simplified model of the converter can be
represented by the circuit of the Fig. 2.
The converter system is designed with following parameters,
which is representative of the experimental system:
Vin = 325V
Io = 0200mA
n = 15
Cs = 30nF
Ls = 24.3H

Vo = 23kV62.5kV
Po = 05kW
fs = 250kHz.500kHz
Cp = 12nF
Co1, Co2 = 0.5F

The equivalent output voltage, output current and output

resistance, referred to the primary side of transformer, are
respectively: Vo = Vo /2n, Io = Io 2n and Ro = Ro/4n2 where n is
the transformer turns ratio.
III.

THEORETICAL ANALYSIS

A. Nonlinear Model of the Series-Parallel Resonant

Converter
State space averaging has been demonstrated to be an effective
method for analysis and control design in PWM switching power
converters. However the method cannot be applied to a wide range
of power circuits that includes the resonant type converters. The
basic limitation in resonant converters is that these circuits have
state variables with a predominantly oscillatory behavior. The
generalized averaging method overcomes the limitations of the
state-space averaging method because it can describe arbitrary
types of waveforms [2]. The method is based on the fact that the
waveform x() can be approximated in the interval (t-T, t] with a
Fourier series representation of the form

x(t T + s ) = x k (t )e jk s ( t T + s )

(1)

where the sum is over all integers k, s=2/T, s (0, T ], and xk(t)
are complex Fourier coefficients.
The time derivative of the kth coefficient is computed to be
d
d
x k (t ) =
x
dt
dt

(t ) jk s x k (t ).

(2)

It is important to point out that (2) corresponds to the exact

derivative of the kth coefficient only in the case that s is constant.
When the frequency is not constant, (2) is only a good
approximation if s is a slowly varying variable [2].
The generalized averaging method is applied to the seriesparallel resonant converter shown on Fig. 2. Using the notation
given in the figure and choosing Co1 = Co2 = Co, the differential
equations that describe the circuit can be written as:
di (t )
(3)
Ls Ls = vCs (t ) vCp (t ) + v AB (t )
dt
dv (t )
(4)
Cs Cs = iLs (t )
dt
dv (t )
2 vo ' (t )
Co ' o ' = abs (iT )
(5)
dt
Ro '
T
The state vector is given by x(t ) = [i Ls (t ) vCs (t ) vo ' (t )] . At
this point, it is important to notice why the voltage in the parallel
capacitor vcp is not used as a state variable even though Cp is an
energy storing element. In fact the vcp waveform is not
differentiable and its dynamic behavior is completely defined by
the energy stored in Ls and in both output capacitors [3].
Fig. 3 shows the main waveforms of the simplified circuit of
Fig. 2 for a typical operating point. These waveforms are
predominantly sinusoidal so it is assumed that the resonant current
iLs and the voltage in the series capacitor vCs are well approximated
by their first time-varying harmonic i.e. with the fundamental term
of the Fourier series. The output voltage vo is a DC voltage with
slow dynamical behavior so it is assumed that it can be
approximated with sufficient accuracy with the DC coefficient of
the complex Fourier series. The nonlinear model for the dynamics
of the local harmonics, assuming that the switching frequency
varies slowly from one switching period to the next, is given by

d I Ls

dt

d VCs

dt

d Vo '

= js I Ls 1 +

1
VCs 1 VCp
Ls

= js VCs 1 +

1
I Ls
Cs

+ VAB

(6)
(7)

2 Vo ' 0
1
abs (iT )1

Co '
Ro '

(8)

Vin
[sin( D) + j (cos( D) 1)]

(9)

dt
where
VAB

 sC p Vo '

0
= cos 1
1
2 I Ls

abs (iT )1 =

2 I Ls

[1 cos() ]

Vin

vAB

t0

-Vin
Ls

t2 t
3

t1

t4 t5

t6

t7 t8

(10)

dx1
x
x
V
= s x2 3 5 + in sin( D)
dt
Ls Ls Ls

(18)

(11)

dx2
x
x
V
= s x1 4 6 + in [ cos( D) 1]
dt
Ls Ls Ls

(19)

dx3
x
= s x4 + 1
dt
Cs

(20)

dx4
x
= s x3 + 2
dt
Cs

(21)

2
2
dx7 2 x1 + x2
2 x7
=
[1 cos()]
dt
Ro ' Co '
Co '

(22)

t9 t10

iLs

t
iT

Ls

vCp

Vo
4n
0

vCs

^
VCs
0

Fig. 3: Time behavior of the characteristic voltages and currents of a seriesparallel resonant converter within a pulse period (t0t10).

The new state variables are: ILs1, VCs1 and Vo0. They are
complex Fourier coefficients that can be written as a fifth order
state space model with real variables: ILs1 = x1 + jx2, VCs1 = x3 +
jx4 and Vo0 = x7. The voltage on the parallel capacitor can be also
written as a function of real variables: VCp1= x5 + jx6 [3].
Eq. (10) and (11) can be rewritten as:
sC p x7

= cos 1
1
2 x2 + x2

1
2

abs (iT )1 =

2 x + x
2
1

2
2

u1 = s

(23)

u2 = D

(24)

y1 = x7

y2 = 2 x + x

(26)

2
4

y3 = 2 x + x

(27)

2
3

Output Voltage

800

600

400

200

(12)
0

[1 cos()]

50

100

150

Resonant Inductor Current

40

20

x5 =

1
[ x1 + x2 ]
sC p

(14)

x6 =

1
[ x2 x1 ]
sC p

(15)

-20

where

Time[ms]

(a)

(13)

As already stated, the voltage on the parallel capacitor is not

considered as a state variable, so one needs to express this voltage
as a function of the state variables. x5 and x6 are expressed as a
function of the existing state variables x1 and x2.

-40

1
= + sin(2)
2
= sin 2 ()

(16)
(17)

The new state vector is given by x = [ x1 x2 x3 x4 x7 ]

the model with real variables is

(25)
2
2

2
1

vo

Vo'

The switching frequency s and duty cycle D are defined as

the control inputs u1 and u2, respectively. The output voltage, the
amplitude of the resonant current and the amplitude of the voltage
in the series-capacitor are defined as the outputs y1, y2 and y3,
respectively.

and

50

Time[ms]

100

150

(b)
Fig. 4: Comparison of the amplitude of the model variables with the
simulated waveforms: (a) output voltage Vo (100V/div); (b) resonant
current iLs (10A/div).

Based on the generalized averaging model, digital simulations

using MATLAB/SIMULINK are performed. The parameters used
for the simulation are: Vo = 767V, Po = 4.6kW, D = 0.752, fs =
253kHz and Ro = 128. The simulation compares the magnitude
of the complex coefficients obtained with the averaging method
with the amplitude of the waveforms obtained simulating the
switching circuit. The results for the resonant current and the
equivalent output voltage referred to the primary side of the
transformer are shown in Fig. 4. The results show that the
correspondence between the waveforms obtained with the model
and with the simulated circuit is excellent. Therefore the large
signal model based on generalized averaging method is a good
starting point to derive the small-signal model of the converter.

B. Small- Signal Model of the Series-Parallel Resonant

Converter
In order to find the small-signal model for the series-parallel
resonant converter with capacitive output filter the first step is to
obtain a steady state solution for the system. This can be done
setting the derivatives of Eq. (18)-(22) equal to zero. Furthermore
the steady-state value of the angle is taken from [1]. The
equations that represent the steady-state solution are given by
2
ss = 2 tan 1
(28)
sC p R o '
1
ss = ss + sin(2 ss )
2
ss = sin 2 (ss )

Cp
sss C p Vin
ss
ss + sss 2 Ls C p

x1ss =

x2ss

+ K sin( D)
M
K M 2 sin( D)

=
+ [ cos( D) 1]

(1 + M 2 ) M

x3ss =

x2ss

x5ss =
x6ss =
x7 ss =

23kV/4.6kW
46kV/5kW
62.5kV/5kW
62.5kV/500W
46kV/2.5kW

80
70

(33)
(34)

60
50
40
30
20
10

(35)

0
-10
360

(36)

sss Cs

x4ss =

90

(32)

ss

Magnitude (dB)

M=

x2ss

(31)

Cs

K=

(30)

(41)

G
where A, B, C and D are matrices that describe the system, x is the
G
JG
state vector, u is the input vector, y is the output vector and
means the small changes of the respective parameter.
It is important to point out that steady state solution is
dependent on the operating point of the system (desired output
voltage and load). Thus, there is one steady state solution for each
operating point of the system. This makes the analytical
linearization of the system complex. The complete set of equations
is presented in the Appendix of this paper. However, if one is
interested in a specific operation point then the numerical solutions
are easily computed with software such as MATLAB. Therefore
this procedure can be used to calculate transfer functions for
desired operating points in a fast and reliable way.
With the linear system formed by the matrices A, B, C, D one
is in position to obtain the six transfer functions that relate the
control input u1 (switching frequency) or u2 (duty cycle, D) to the
outputs y1 (output voltage), y2 (amplitude of the resonant current)
or y3, (amplitude of the voltage in the series capacitor). The
transfer functions that relate the inputs u1, u2 with the output y1, are
particularly interesting because the main control purpose is to
adjust the output voltage to any desired reference value for all kind
of load changes. The switching frequency of the converter is
automatically adjusted to ensure zero-current switching of one
bridge leg where the output power is controlled by duty cycle
variation. Therefore, the transfer function that is essential for the
controller design is the transfer function that relates duty cycle
with output voltage G(s) = Vo / D = y1 / u2.

x1ss

(37)

sss Cs

270

Phase (deg)

(29)

x = Ax + B u

y = C x + D u

180
90
0

x1 ss + x2ss ss
sss C p ss

(38)

-90
-180

x2 ss x1ss ss
sss C p ss
Ro ' x1ss 2 + x2ss 2

[1 cos(ss )]

(39)

10

10

10

10

10

Frequency (Hz)

Fig. 5: Bode plots of G(s) for different operating points of the system.

(40)

After obtaining the steady-state solution, the model can be

linearized around the steady state in order to obtain the smallsignal transfer functions from any desired input to any desired
output. After the linearization one obtains a linearized model of
the type given by Eq. (41).

IV. SIMULATION RESULTS

Fig. 5 presented the Bode magnitude and phase plots of the
output voltage for small changes in duty cycle. They represent the
transfer function G(s) for different load conditions from full power
to very low power. The five different output voltage / output power
operating points are: 23kV/4.6kW, 46kV/5kW, 62.5kV/5kW,
62.5kV/0.5kW and 46kV/2.5kW. These points are arbitrarily

10

chosen just in order to show how the transfer function changes

within the operating range of the converter.
One can see that the DC gain of the transfer function G(s)
changes significantly with the load. The phase plot shows that for
some load conditions the transfer function has two non minimum
phase zeros, which indicates that the system is difficult to control.
This characteristic is evident comparing phase transitions for the
same output voltage, 62.5kV, and different output power, 5kW and
500W. For 5kW the system is non minimum phase and for 500W
is minimum phase. Thus a very robust control technique is needed
in order to overcome this difficulty. The peaks in the magnitude
correspond to summation of the series resonant frequency fo and
the switching frequency fs.
Output Voltage [V]
900
850
800
750

occurs. This represents the step response when the reference

voltage (referred to the primary side of the transformer) is held
constant at 767V and the output power is reduced from 4.6kW to
2.3kW.
From the results shown in Fig. 6, one can conclude that the
linearized model agrees well with the dynamics of the simulated
converter. Therefore the calculation method is valid and is a very
helpful tool for the control design.
V. EXPERIMENTAL RESULTS
In order to verify if the results obtained theoretically and with
digital simulations are reproducible in practice, a 5kW prototype is
built. Since the objective of the laboratory tests is only to verify
the validity of the small-signal representation, one does not have to
construct the complete high-voltage converter. It is sufficient to
build the equivalent circuit shown in Fig. 2 with all quantities
referred to the primary side of the transformer. Thus, the prototype
is constructed without high-voltage transformer. The parasitic
elements (leakage inductance and stray capacitance), which are
necessary to build the resonant circuit, are replaced with external
elements. The photograph of the constructed prototype of the
equivalent system is shown in Fig. 7.

700
650
600
500

19.5 cm

700

900
1100
Time[ms]
(a)

1300

1500

16 cm

Output Voltage [V]

850

800

750

700
700

900

1100

1300
1500
1700 1900
Time[ms]
(b)
Fig. 6: Comparison between step response of the linearized model and of
the circuit (a) with a step change in the value of reference voltage and (b)
with a step change in the value of the load resistance. Scale: output voltage
Vo (50V/div).

In order to prove that the small-signal model well

approximates the converter, simulations using MATLAB were
undertaken for various load and reference output voltage steps.
Fig. 6(a) shows a comparison between the simulated switching
circuit with a constant load and the linearized model. It represents
the closed loop response to a step change in the reference voltage
Voref from 700V to 800V using a PI controller with fixed
parameters. Fig. 6(b) shows a comparison between the simulated
switching circuit and the linearized model when a load change

Fig. 7: Photograph of the 5kW prototype of the series-parallel resonant

converter with capacitive output filter.

Fig. 8 shows the comparison between the results obtained with

simulation and from the laboratory tests without the high-voltage
transformer. This represents the closed loop response of the system
when a step change in the reference voltage occurs. At time of
0.1ms the reference voltage is changed from 410V to 750V. In the
figure, one can see that the rise time, which is about 100s, is the
same for both simulated and experimental waveforms. The small
differences between both waveforms are due to approximations
used in the development of the model and to the unmodeled nonidealities of the real circuit. In the practical implementation the
reference voltage passes through a first order proportional/lag
element before it is compared with the measured voltage.

Therefore the measured output voltage is compared with a

reference voltage that has an exponential rise rather than a step
change. Since the controller is very fast, the controller sets the duty
cycle immediately to the maximum value when a step occurs and
the output voltage is, for a short time, higher than the reference
voltage. The controller then reduces the duty cycle to a minimum
value and this transient behavior is responsible for the plateau that
occurs from time 0.2ms to 0.3ms in the experimental waveform
before the output voltage reaches its final value. Even considering
this small difference between both waveforms, there is a good
agreement between the simulation results using the proposed
small-signal model and the experimental result.

APPENDIX
Matrix A is a 5x5 system matrix and it gives the relation
between the derivative of states and the states.
a11 a12
a
21 a22
A = a31 0

0 a42
a
51 a52

a13
0
0
a43
0

a15
a25
0

0
a55

0
a24
a34
0
0

where
900

I1ss = x1ss 2 + x2ss 2

800

x7ss s C p

(A.1)

G1 =

700

4 Ls I1ss

x7ss

Ls I1ss

x7ss
x7 2 s C p
ss
G2 =

Ls I1
2 Ls I1ss 2
ss

600

(A.2)
(A.3)

800

x7ss 2 s C p tan ss
2
G3 =
+
2 Ls I1ss 4
ss
2
s C p Ls I1ss tan
2
(A.4)

x7ss
1

tan ss +

3
2 Ls I1ss 2
ss
tan

ss
1
+
(A.5)
G4 = s
s C p Ls s C p Ls

700

a11 = G1 +

600

400

x1 x2
1

a12 = ss 2 ss + tan ss
2 2
I1ss

1
a13 =
Ls

300

a15 =

500

400
300
0

0.1

0.2

0.3

0.4

0.5 0.6
Time[ms]

0.7

0.8

0.9

(a)
900

500

0.1

0.2

0.3

0.4

0.5
0.6
0.7 0.8
0.9
1
Time[ms]
(b)
Fig. 8: Closed-loop response of the output voltage for a step change in the
reference from 410V to 750V: (a) Simulation result using the proposed
model (b) Experimental result. Scale: output voltage Vo (100V/div), Time
(100s/div).

VI. CONCLUSIONS
This paper presents an accurate method, based on first harmonic
analysis, to derive the small-signal model of the series-parallel
resonant DC-DC converter with capacitive output filter. The model
provides a fast and reliable way to obtain desired transfer functions
of the system for each operating point. For this reason, the smallsignal model simplifies the controller design task for resonant
converters by removing the need of complex mathematical
analysis and saves the time required for measurements in the trial
and error design method.

x1ss
x7ss

x1ss 2
I1ss 2

G2 + x1ss x2ss G3

G2

x2ss I1ss 2
x7ss

2
G2 + x2ss G3 + G4

G3

(A.6)

(A.7)
(A.8)
(A.9)

x1 x2
1

a21 = ss 2 ss tan ss G2 x1ss 2 G3 G4
2
2
I1ss

(A.10)

a 22 = G1 + x2ss 2 G2 x1ss x2ss G3

(A.11)

a 24 =
a25 =

1
Ls
x2ss
x7ss

(A.12)
G2 +

x1ss I1ss 2
x7ss

G3

(A.13)

1
Cs

(A.14)

a34 = s

(A.15)

1
Cs

(A.16)

a31 =

a 42 =

a 43 = s
4 x1ss

a51 =

(A.18)

Co ' I1ss

a52 =
a55

(A.17)

4 x2ss

(A.19)

C o ' I1ss

1
=
Co'

s C p
2

+R
o'

(A.20)

B is the input matrix and it relates the derivative of the states

with the system inputs. The two inputs, D and s are used resulting
in a 5x2 matrix for B.

2s C p Ls tan ss
2

2 s Ls I1ss tan ss
2

(A.25)

b41 = x3ss

(A.26)

b22 =

Vin
sin( D)
Ls

I1ss
2 x2ss

(A.32)

I1ss
2 x3ss

(A.33)

V1ss
2 x4ss

(A.34)

V1ss

REFERENCES

[2]

[3]

b31 = x4 ss

Vin
cos( D)
Ls

c22 =

x7 ss

(A.24)

b12 =

(A.31)

[1]

b21 = x2ss H 1 x1ss H 2

Co '

0
0

2 x1ss

c21 =

(A.21)

(A.22)
(A.23)

b51 =

0
0
c34

Finally, matrix D is called the direct transmission matrix and it

represents the elements that transmit an input directly to the output
of the system. For our case D is a 3x2 zero matrix.

b11 = x1ss H 1 + x2ss H 2

x7ss C p

0
0
c33

(A.30)

c34 =

ss
1

2s C p Ls 2s C p Ls

x7ss 2 C p tan ss
2 +

4 Ls I1ss 2

0
c22
0

V1ss = x3ss 2 + x4ss 2

c33 =

b11 b12

b21 b22
B = b31 0

b41 0

b
51 0
where
x7ss 2 C p
H1 =
4 Ls I1ss 2
H2 = 1 +

C = c21
0
where

[4]
[5]

(A.27)
(A.28)

[6]

(A.29)

The matrix that relates the states with the output is the C
matrix. There are three outputs and five states thus C is a 3x5
matrix.

[7]

F.S. Cavalcante and J.W. Kolar, Design of a 5kW High Output

Voltage Series-Parallel Resonant DC-DC Converter, in Proceedings
of the 34th IEEE Power Electronics Specialists Conference, Acapulco,
Mexico, 2003, vol. 4, pp. 1807 - 1814.
S.R. Sanders, J.M. Noworolski, X.Z. Liu, et al. Generalized
Averaging Method for Power Conversion Circuits, IEEE
Transactions on Power Electronics, vol. 6, no. 2, pp. 251-259, April
1991.
J.A. Martn-Ramos, J. Diaz, A.M. Perna, et al. Large-Signal
Modeling of the PRC-LCC Resonant Topology with a Capacitor as
Output Filter, in Proceedings of the 17th Annual IEEE Applied
Power Electronics Conference and Exposition, Dallas, TX, 2002, vol.
2, pp. 1120 - 1126.
V. Vorprian, Approximate Small-Signal Analysis of the Series and
the Parallel Resonant Converters, IEEE Transactions on Power
Electronics, vol. 4, no. 1, pp. 15-24, January 1989.
E.X. Yang, B. Choi, F.C. Lee, et al. Dynamic Analysis and Control
Design of LCC Resonant Converter, in Proceedings of the 23rd
Power Electronics Specialists Conference, Toledo, Spain, 1992, vol.
1, pp. 362 369.
A.M. Stankovic, D.J. Perreault, K. Sato, et al. Analysis and
Experimentation with Dissipative Nonlinear Controllers for Series
Resonant DC/DC Converters, in Proceedings of the 28th Power
Electronics Specialists Conference, St. Louis, MI, 1997, vol. 1, pp.
679 685.
G. Garcia Soto, J. Gaysse, G.W. Baptiste, Variable Sampling Time
Serial-Resonant Current Converter Control for a High-Voltage X-ray
Tube Application, in Proceedings of the 10th European Power
Quality Conference, Nuremberg, Germany, 2004, pp. 972 - 977.