Professional Documents
Culture Documents
Current Sensor Fault Detection by Bilinear Observer For A Doubly Fed Induction Generator
Current Sensor Fault Detection by Bilinear Observer For A Doubly Fed Induction Generator
Current Sensor Fault Detection by Bilinear Observer For A Doubly Fed Induction Generator
+
JJG
G JG
(2)
R
R R
R
R R R
d
U = R I
dt
+
JJG
G JG
(3)
S
S S S
R
R R R R m
d
U = R I jp
dt
+
JJG
G JJG JG
(4)
A
A A A S
S S S S A
A
A A A R
R R R R A m
d
U = R I j
dt
d
U = R I j( p )
dt
+ +
+ +
JJG
G JJG JG
JJG
G JJG JG
(5)
Using (6) and splitting the resulting equations into real and
imaginary part, one obtains:
A A A
S S R S h
A A A
R S R h R
= L I L I
= L I L I
+
+
JJG G G
JJG G G
(6)
A A
A A A A sd rd
sd s sd A s sq A h rq s h
A A
sq rq A A A A
sq s sq A s sd A h rd s h
A A
A A A A sd rd
rd r rd A m h sq A m r rq h r
A A A A
rq r rq A m h sd A m r rd
dI dI
U = R I L I L I L L
dt dt
dI dI
U = R I L I L I L L
dt dt
dI dI
U = R I ( p )L I ( p )L I L L
dt dt
U = R I ( p )L I ( p )L I
+ +
+ + + +
+ +
+ +
A A
sq rq
h r
dI dI
L L
dt dt
+ +
Solving for the differential terms and defining
2
h
S R
L
1
L L
=
one obtains four differential equations, which can be written
in the known state space equation:
dx
Ax Bu; y Cx
dt
= + =
(7)
The input vector u is defined by the real and imaginary
parts of the stator and rotor voltages (8), the state vector x
comprises of the real and imaginary parts of the stator and
rotor currents (9).
T
sd sq rd rq
u U U U U ( =
(8)
T
Sd Sq Rd Rq
x I I I I ( =
(9)
The system matrix A, the input matrix B and the output
matrix C are given in (10), (11), and (12).
2
S h h R h
A m m
S S R S R S
2
S h h h R
A m m
S R S S S R
h S h R
m A m
R S R R
h S h R
m A m
R R S R
R L L R L
p p
L L L L L L
R L L L R
( p ) p
L L L L L L
A =
L R L R 1
p p
L L L L
L R L R 1
p ( p )
L L L L
(
+
(
(
(
+ (
(
(
(
(
(
(
(
h
S R S
h
S R S
h
R S R
h
R S R
L 1
0 0
L L L
L 1
0 0
L L L
B =
L 1
0 0
L L L
L 1
0 0
L L L
(
(
(
(
(
(
(
(
(
(
(
(10)
(11)
1 0 0 0
0 1 0 0
C =
0 0 1 0
0 0 0 1
(
(
(
(
(
(12)
Note that the system matrix A still contains the rotational
speed
A
of the arbitrary reference frame and therefore is a
general description for all reference frames. The system
matrix for the stator-fixed reference frame is obtained by
setting
A
=0. Likewise, the system matrix for the rotor-fixed
reference frame is obtained by setting
A
=
m.
When using a reference frame fixed to the stator, the
indices dq are usually replaced by .
As can be seen from (10), A contains the rotor speed
m
as
a variable input parameter, which means that the system is
non-linear. Therefore, A is split up into A
0
which represents
the elements independent of
m
, and N, which contains the
elements dependent of
m
[9] (13).
0 m
A A N = +
(13)
For the stator-fixed system, N
S
shall be used. A state space
system like this is called bilinear, because the non-linear
input
m
can be isolated. Then A
0
is linear, and N
m
is linear
as well for constant
m
, and linearly dependent on
m
.
2
h h
S R S
2
h h
S R S
S
h
R
h
R
L L
0 0
L L L
L L
0 0
L L L
N =
L 1
0 0
L
L 1
0 0
L
(
(
(
(
(
(
(
(
(
(
(
(
(14)
S h R
S S R
S h R
S S R
0
h S R
R S R
h S R
R S R
R L R
0 0
L L L
R L R
0 0
L L L
A =
L R R
0 0
L L L
L R R
0 0
L L L
(
(
(
(
(
(
(
(
(
(
(
(15)
Modelling of Stator and Rotor Current Sensor Faults
In case of sensor faults, some states are either not
measurable or shall not be used for feedback because of false
measurements. This is described by introducing the effective
output matrix C
eff
according to (16). If all sensors are used for
feedback, F becomes the identity matrix.
1
2
eff
3
4
f 0 0 0
0 f 0 0
C C F C
0 0 f 0
0 0 0 f
(
(
(
= =
(
(
(16)
( )
i
1 sensor fedback
f i 1, 2, 3, 4
0 sensor not fedback
(17)
Placement of the Observer Eigenvalues
According to [6] and [7], Luenberger observers can be used
for residual generation. Using Luenberger observers, the
eigenvalues of the closed loop observer can be placed by
choosing an appropriate feedback matrix L.
eff
eff
eff
d
x Ax Bu
dt
d
x Ax Bu LC (x x)
dt
d d d
e x x
dt dt dt
d
e Ax Bu (Ax Bu LC (x x))
dt
d
e (A LC )e
dt
= +
= + +
=
= + + +
=
(18)
By doing so, the dynamical behaviour of the estimation
error e can be defined. For linear systems, the error dynamics
are described by (A-LC
eff
) (18), where x and x are the
estimated and measured state vectors, respectively, and e is
the error vector.
Similar to (13), L
0
and M are defined by (19), leading to
(20). Figure 3 shows the structure of this bilinear observer.
u
y
A
0
C B
1
s
x
N
S
+
+
L
0
x
.
+
+
-
+
mech
A
0
C B
1
s
N
S
+
+
+
+
.
.
.
+
+
y
x x
.
Doubly Fed Induction Machine
Model
u
Residual Generation and
Feedback
Figure 3: Topology of bilinear observer
0 m
L L M = +
(19)
0 0 eff S eff m
d
e (A L C (N MC ) )e
dt
= +
(20)
As can be seen, the error dynamics are again defined by
A
0
-L
0
C
eff
if M is defined by MC
eff
=N. This is only possible if
C
eff
is invertible. This is the case if all sensors are used for
feedback. In this case (21) and (22) will define the
eigenvalues
i
of the error dynamics independently of the
rotor speed
m
.
[ ]
T
0 0 4 1 2 3 4
L A I = +
(21)
M N =
(22)
Observer without complete set of sensors
Observers are useful for observing states that shall not or
cannot be measured. Therefore, the case of an observer
without a complete set of sensors is more realistic. In this
case, the effective output matrix C
eff
(16) is not invertible.
Therefore, MC
eff
=N is not possible, and placing the
eigenvalues becomes a non-linear function of
m
.
In this case, the Matlab-function place is used to place
the eigenvalues according to the algorithm described in [12].
The problem of non-linear observer feedback is not
addressed in this paper.
III. IMPLEMENTATION OF THE OBSERVER
STRUCTURES
Referring the Measured Signals to the Stator
The observer is fed with the stator and rotor voltages
referred to the stator. Residuals are computed with additional
input of stator and rotor current also referred to the stator.
Therefore, the rotor signals that are only available in the rotor
system need to be processed in order to obtain the referred
signals. This requires a change in frequency from rotor to
stator frequency. This is achieved by implementing a vector
rotation with the angle of the rotor position. Usually a change
in amplitude is necessary because of the transformation ratio
of the DFIG. The transformation ratio is derived from the
number of turns of stator and rotor winding. Therefore, the
transformation ratio is implemented by a gain factor for the
rotor current and voltage. The measurements need to be
converted into space vectors by concordia transform. The
rotor voltage is not measured. Instead, the commanded rotor
voltage is used, decreased by a factor for dead time
compensation. The matrices F and I
4
-F are used to separate
the residuals into feedback residuals and detection residuals.
Realisation of a Fault-Detection Observer
The structures depicted in figures 4 to 7 are realised in
Matlab-Simulink for simulation and experiment via a
dSpace Board. Integration is approximated by a 500s
Forward-Euler algorithm.
Figure 4: Overview of observer structure
Figure 5: Signal Processing. Referring the rotor signals to the stator system.
Figure 6: Model: Bilinear state space model of the DFIG.
Figure 7: Residual and feedback: Calculation of the error feedback and
residuals for detection
Figure 8: Stator and rotor currents (experiment). Red: observed currents.
Green: Measured currents. Activation of error feedback at t=0s, feedback of
stator and rotor currents
Figure 9: Stator and rotor current residuals (experiment). Activation of error
feedback at t=0s, feedback of stator and rotor currents.
IV. EXPERIMENTAL BEHAVIOUR
No Error Feedback vs. Complete Error Feedback
In case of ideal modelling and ideal knowledge of the plant's
parameters, the state space model itself should follow the
plant without feedback. These assumptions can hardly ever
be met, yet it is important for the model to be as close as
possible to the plant. Therefore the behaviour of the open
loop model and the closed loop observer is presented in
figure 8. Machine data is given in table 1 in appendix VI.
Note that the parameters used in the observer need to be
chosen larger than the identified machine parameters, as
stated in table 2. Figure 8 shows the measured and estimated
stator and rotor currents before and after closing the feedback
loop at t=0s. Without feedback, a large subharmonic can be
noticed in the open loop model estimates, while the
fundamental component is close to the measured
currents.With activated feedback, the closed loop observer
estimates and the measured currents fit very good. Figure 9
shows the stator and rotor residuals
S
and
R
. As can be
seen, the residuals become small with activated feedback.
Figure 10: Stationary stator and rotor currents (experiment). Red: observed
currents. Green: Measured currents. Blue: Residuals. Activation of error
feedback at t=0s, feedback of rotor currents only
Behaviour with Error Feedback of two States
There is no fault detection capability in case all four states
are fed back, because the observer will be affected by any
false current measurement. Feeding back only the measured
rotor currents, the stator currents can be checked for sensor
faults. The behaviour of a two state feedback is documented
in figure 10. It is shown that the subharmonic component is
eliminated in both the fed back rotor currents and in the
stator currents, that are not fed back. The stator current
residuals increase compared to a four-current feedback in
figure 9, but the stator currents residuals are now decoupled
from the stator current sensors. This makes stator current
sensor fault detection possible.
Behaviour of the Observer During Transients
It is important that the currents estimated by the closed loop
observer follow the measured currents not only in steady
state, but also during transients. To demonstrate this, figure
11 shows the behaviour of the observer during a step of the
reference current. The rotor current reference steps up from 4
to 28 A at t=0s. Note that the current depicted in figure 11 is
the referred current and therefore the actual rotor current
divided by the transmission ratio of the DFIG. The
corresponding residuals are shown during the steps. It can be
seen that the residuals are almost not responding during both
steps. This is an important feature, since any fault detection
must not be triggered by transients.
Figures 12 and 13 show the behaviour of an observer using
only the rotor currents as feedback. This time, a step upwards
and downwards is shown. The stator current residuals are
more affected than in the case of feeding back all four
currents. They are, however, still small and the peak during
the transients declines fast.
Figure 11: Step response of stator and rotor currents (experiment). Red:
observed currents. Green: Measured currents. Blue: Residuals. Activation of
error feedback at t=0s, feedback of stator and rotor currents.
Figure 12: Step response of stator and rotor currents (experiment) Red:
observed currents. Green: Measured currents. Step at t=0s and t=2s,
feedback of rotor currents only.
Figure 13: Residuals during the steps. Step at t=0s and t=2s, feedback of
rotor currents only.
Using the Free Stator Current Residuals for Fault Detection
If only two currents are fed back, the other two may be
used for fault detection of the respective current sensors. In
order to show this capability, the observer stator current
inputs are switched to zero to simulate a sensor fault. The
inputs to the current controllers are not affected, though, to
maintain a working control.
Due to the character of the concordia transform (23), a fault
in stator phase U will affect the and component of the
stator current. A fault in phase V will affect the -component
only. Therefore, a fault in stator phase V can be detected by
an increase of the -residual. A fault in phase U will lead to a
change in both the and -residual, where the change in the
b-component will be smaller than the change resulting from a
fault of phase V. A fault of both current sensors will also lead
to an increase of both residuals.
U
V W V U
V U
I I
I I 2 I I
I 1.15 I 0.58 I
3 3
=
+
= = = +
(23)
Figure 14: Stator and rotor current residuals. Current sensor failure in phase
U (blue flag), phase V (red flag), and both phase U and V. Feedback of
rotor currents only.
This is difficult to distinguish from a fault in phase U, since
the magnitude of the residuals is only slightly different. This
behaviour is shown in figure 14. Detection of a current
sensor fault in phase U may be realised by cyclic swapping
of the three phases, so that phase U will only affect a new
residual *. Detection of rotor current sensor fault may be
realised by using an observer in the rotor reference frame.
V. CONCLUSION
A bilinear state observer is used to detect current sensor
faults. The observer may be stabilised by using a feedback of
rotor currents only, therefore leaving the stator current
residuals for detection. The proposed observer may serve two
purposes. Firstly, it serves as a residual generator for sensor
fault diagnosis. It is shown that a significant residual can be
generated for a stator current fault in phase V. Residuals for
other current sensors can be generated as well. Secondly, the
observer supplies an approximated signal for the faulty
measurement. This may be used to maintain a controlled
operation of the drive, when fault tolerant control is to be
realised. Further research in this direction is necessary.
VI. DATA OF EXPERIMENTAL SETUP
Machine VEM
SPER 200 LX4
22 kW Stator: 400 V 41 A
Rotor: 255 V 53 A
Control dSPACE DS1104 250 Mhz
Inverter (DFIG) IGBT 2-level Voltage Source Inverter
Converter (Grid) 3-phase Diode Rectifier
Table 1: Data of experimental setup
Parameters Identified Used in Observer
Lh 48 mH 48 mH
Ls 49.1 mH 51.1 mH
Lr 49.1 mH 51.6 mH
Rs
100 m 200 m
Rr
250 m 300 m
Transmission Ratio 1.5 1.5
Table 2: Doubly Fed Induction Machine Parameters
VII. ACKNOWLEDGEMENT
This work was funded by the Deutsche
Forschungsgemeinschaft (German Research Foundation).
VIII. REFERENCES
[1] A. Hansen, F. Iov, P. Srensen, F. Blaabjerg. Overall control strategy of
variable speed doubly-fed induction generator wind turbine. Nordic
wind power conference 2004, CD-ROM, Gteborg, Sweden, 2004.
[2] A. Bocquel, J. Janning: Analysis of a 300 MW Variable Speed Drive for
Pump-Storage Plant Applications, EPE'05, CD-ROM Paper, Dresden,
Germany, 2005.
[3] H. Akagi, H. Sato: Control and Performance of a Doubly-Fed Induction
Machine Intended for a Flywheel Energy Storage System, IEEE
Transactions on Power Electronics, Vol. 17, No.1, pp. 109-116, Jan
2002.
[4] R. Burschen: Shaft Generator Plants with Slipring Induction Machines,
Treffen Schiffsbautechnische Gesellschaft e. V., pp. 18-21, 11, 1987,
Hamburg.
[5] F. Khatounian, E. Monmasson, F. Berthereau, E. Delaleau, J. Louis:
Control of a Doubly Fed Induction Generator for Aircraft Application,
IECON'03, pp. 2711-2716, Roanoke, USA, 2003
[6] R. Patton, P. Frank, R. Clark: Issues of Fault Diagnosis for Dynamic
Systems. Springer-Verlag, Berlin, 2000.
[7] T. Steffen: Control reconfiguration of dynamical systems : linear
approaches and structural tests. Springer Verlag, Berlin, 2005
[8] R. Lorenz: Observers and state filters in drives and power electronics.
Journal of Electrical Engineering, Vol. 2, 2002.
[9] S. Bennett: Model Based Methods for Sensor Fault-Tolerant Control of
Rail Vehicle Traction, PhD-Thesis, University of Hull, 1998.
[10] S. Bennett, R. Patton, S. Daley, D. Newton: Model Based Intermittent
Fault Tolerance in an Induction Motor Drive, Symposium on Control,
Optimization and Supervision, Vol.1, pp 678-83, Lille, 1996.
[11] C. Thybo: Fault-tolerant Control of Inverter Fed Induction Motor
Drives. PhD Thesis, University of Aalborg, 1999.
[12] Kautsky, J. and N.K. Nichols, Robust Pole Assignment in Linear State
Feedback, Int. J. Control, 41 (1985), pp. 1129-1155