Sunteți pe pagina 1din 6

Data-driven recongurability evaluation of linear

systems: Simulation application in a DC motor


B.M. Gonzalez-Contreras, E. Juarez-Guerra D. Sauter, D. Theilliol
Universidad Autonoma de Tlaxcala, Centre de Recherche en Automatique de Nancy (CRAN)
Dept. of Engineering, School of Electronics (ISE) CNRS UMR 7039, Nancy Universite,
CP 90300 Apizaco, Tlaxcala, Mexico BP 239 - 54506 Vanduvre CEDEX-France.
Phone: +52 241 4172544 Phone: +33 383 684 465, 70 239
Fax: +52 241 4172543 Fax: +33 383 684 462
e-mail: brianmgc@ieee.org e-mail: dominique.sauter@cran.uhp-nancy.fr
didier.theilliol@cran.uhp-nancy.fr

AbstractTolerance to faults of control systems has been indicated in [9]. However, these measures must be done in the
dened through the control recongurability concept, which is pre-design phase, i.e., before the system begins to operate, and
a property dening the capacity of the control system to allow therefore they are evaluated off-line. For on-line purposes the
restoration of performance in the presence of faults. For its eval-
uation the controllability gramian is required. This evaluation is proposed measures could be more useful for applications such
usually done before control system be implemented, i.e., off-line. as redesign, remodeling [7] and supervision of a process [10].
Instead, the on-line evaluation of the control recongurability, It is worth pointing out that these measures have a common
from input/output data when the system operates under given base: the gramian concept. Moore [6] stated that the structure
conditions (presence of faults), could be useful for fault tolerance of the controllability gramian, or its singular values and vec-
redesign, analysis and control system quality assessment. The
contribution of the paper is to propose a data-driven way tors, characterizes how fast and in which directions the asso-
of calculating on-line the controllability gramian in order to ciated system dissipates energy. This is viewed as an ellipsoid
evaluate an index based on the control recongurability. We characterizing those states reachable by a unit energy input.
propose to use an identication algorithm to compute the control Thus, it tells something about the energy distribution that
recongurability of the system under partial loss of effectiveness goes through the state variables from the input excitation. In
type faults. Simulation of a direct-current motor application
example illustrating the approach is presented. addition, the existence of an invertible nite gramian ensures
the controllability of the system [11]. Consequently a treated
I. I NTRODUCTION system is called control recongurable if the controllability
Development of fault tolerant control (FTC) systems has property of the system is kept in a faulty situation [4].
been motivated by different goals depending on the application From identication point of view, a similar interpretation
under consideration, principally in ight control [1] and then of the gramian is employed. Different lengths of the ellipsoid
in industrial processes [2]. The basic idea is to achieve axes indicate that some directions are more excited from the
graceful degradation in performance (at worst) in the event perspective of an external input excitation [12], [13]. Hence
of a systems fault associated with an actuator, sensor or that recongurability can be evaluated from the input/output
component. Hence the research topic has been very active data provided from the system already developed.
in the recent years [3], [4]. For analysis, evaluation and In this paper, we propose an index based on the control
synthesis in relation to FTC system design, some measures recongurability (Section III) to establish the admissible op-
of fault tolerance have been proposed. The quotient between eration conditions of the treated system under possible partial
the controllability gramian norm in the nominal situation and loss of effectiveness type faults. Next, we use the Eigensystem
the faulty one was proposed in [5] as a measure of how Realization Algorithm (ERA) to compute the controllability
well the system still can be excited after a fault occurrence. gramian (Section IV) in an indirect way from input/output
Using the smallest second order modes of a system [6], the data obtained on-line (data-driven).
recongurability concept was proposed in [7]. This notion, The index, in terms of the systems closed-loop output with
currently used in the fault tolerant control area, establishes respect to known inputs, can be useful for practical purposes
the potentiality of a process to maintain a certain performance in the supervision, analysis and redesign of the FTC system
under faulty circumstances. In a similar interpretation to [5], already established for a process. In this work, an ideal fault
the system recongurability was latter presented as the quality detection and isolation (FDI) module is assumed in order
of the control system under given faulty conditions [8]. to provide the fault characteristics after fault occurrence. A
The above mentioned measures representing the control simulation example (Section V) of a direct current (DC) motor
recongurability, i.e., the control system quality/capacity to in separately-excited conguration is used for illustration of
respond to faults, can be viewed as a recongurability-based using the index in off-line and on-line fashion. The following
performance or as an intrinsic recongurability property, as section presents the notations and the general linear model

978-1-4244-6391-6/10/$26.00 2010 IEEE 125


assumed in the remainder of the paper. where i , i = 1, . . . , n, indicates the ith eigenvalue of
the associated matrix.
II. M ODEL AND PRELIMINARIES
Using the form 1 or 2, i.e., equations (5) or (6), the
A process described by a linear discrete state-space repre-
criterion proposed in [7] to know how well the model remains
sentation with dependence of a sampling period h is:
controllable from the input to the output, is computed as
x(k + 1) = A x(k) + B u(k) follows:
(1)
y(k) = C x(k), % = min {i } , i = 1, . . . , n, (7)
i
where vectors x(k) Rn , u(k) Rr , y(k) Rm , and where % is the control recongurability.
matrices A Rnn , B Rnr , C Rmn . If the system is Consider now the criterion proposed in [8]:
unstable, then a state feedback is assumed under the following { (( )1 )} [ { }]1
form: = max i Wcb = min i (Wcb ) , (8)
i i
u(k) = Kx(k), (2)
where represents the maximal energy consumed by actuators
with K Rrn in order to become (1) stable. Therefore A (the alternative notion of control recongurability). Note that
or Acl = A + BK (stables) can be used through the following under a balanced representation, criterion (7) is the inverse of
presentation. The triplet (A, B, C) represents the system (1) criterion (8).
and the pairs (A, B), (A, C) are supposed to be controllable The advantage of the former is that in cases of almost
and observable, respectively. outages ( 1) of sensors or actuators, the value is bounded
The faults considered in this work are loss of effective- below by zero, and under no faulty condition the value is
ness [[7] represented
]T in terms of the effectiveness vector always bounded above by the minimum Hankel singular value.
= 1 , . . . r , where i is the effectiveness factor [14] Hereafter the use of the controllability gramian should be
i = 1, . . . , r, with 1 i < 0. This loss of effectiveness considered, although the observability gramian could also be
can affect both actuators and sensors. Under this framework, if chosen taken into account that a balanced representation has
i = 0 the actuator (sensor) is normal, if i = 1 the actuator been adopted.
(sensor) is lost. Throughout this work limit values = 1 are Consider the criterion (7) in order to propose an index based
not considered. on recongurability Q% , as follows:
The following notation will also be useful. Let %f
Lyap(X1 , X2 ) denote the positive denite solution Q Q% = (100%), (9)
%n
of the Lyapunov equation (3) affecting the pair of compatible
where %n is the value in the nominal case (fault free), %f is
matrices (X1 , X2 ):
the value for % ranging between the nominal value %n and
X1 QX1T + X2 X2T = Q, (3) 0 (total effectiveness loss), namely, the % value under faulty
T conditions. Note that Q% values will fall in the range [100, 0]%
where represents the transpose of a matrix. Then the
due to the normalization, where 100% is equivalent to the best
controllability and observability gramians are respectively
value and 0 the lowest value.
Wc = Lyap(A, B) and Wo = Lyap(AT , C T ).
The following denition of admissibility [8] is considered
III. A N INDEX BASED ON RECONFIGURABILITY in order to establish limits on the overall system operation
The two principal types of evaluation of the control recon- taking into account the index (9).
gurability (in open or closed-loop) are reviewed. In the rst Denition 1. A solution of the control problem of the faulty
one, one of the following values is required: system is admissible with respect to a control objective if
1) From the original system representation nd a regular Q% > Qad , (10)
matrix Tb (use the algorithm presented in [11] for
example), leading to a balanced representation, where where Qad is a predened threshold, which represents the
system matrices are given by maximal loss of efciency that can be admitted when a control
solution is used, provided that this solution achieves the
Ab = Tb ATb1 , B b = Tb B, C b = CTb1 , (4) control objective under a faulty situation.
then the gramians w.r.t. these matrices are equal and Based on Qad , admissible values for the solutions or ad-
diagonal: missible operation conditions with this value are established in
order to limit the worst system operation under the occurrence
Wcb = Wob = diag(1 , . . . , n ), (5)
and presence of faults. This value is settled in order to establish
where each i , i = 1, . . . , n, is known as a Hankel fault accommodation, control reconguration or stopping oper-
singular value. ation. The choice may consider to take into account attainable
2) Using the observability and controllability gramians power consumption, actuator saturation [15] or observability
from the original system compute the second order for detectability/isolability [16].
modes i2 : Computation of Q% can be done off-line or on-line. The
i2 = i (Wc Wo ), (6) latter depends on the evaluation of controllability gramian

126
from input/output data provided by the actual system mea- Now a solution for the observer Markov parameters is obtained
sures. In the following, data-driven computation of gramians through least squares (+ denoting pseudo-inverse of a matrix):
is proposed.
Y = V + , (17)
IV. DATA - DRIVEN COMPUTATION
provided that V V T be full rank. System MP are obtained by
The Eigensystem Realization Algorithm (ERA) [12] is partition of the observer MP as follows
proposed here to nd a triplet (A, B, C) representing the [ ]
system (1). Different to other realization theory based methods Y = 0 C B C AB C Ap1 B
[ ] (18)
for state space identication, as that of subspace methods = Y0 Y1 Y2 Yp ,
[17], ERA employs the largest singular values representing
the input/output relationships, which also characterize the
thus one ] directly Y0 = 0 and from A = A + GC and
[ obtains
B = B G one has:
system recongurability. The Markov parameters (MP) are
used in order to capture the system dynamics and to obtain a Yi =C Ai1 B
[ ]
representation given by the excited states. However, obtaining = C(A + GC)i1 B C(A + GC)i1 G (19)
MP is not directly done. A rst step is required, where the [ ]
MP of an observer are computed and then the system MP. = Yi(1) Y (2)
i
, i = 1, 2, . . . , p.

A. Markov parameters from input/output data Now the system MP are extracted from this ordering by using:
The identication technique known as Observer/Kalman
Filter Identication (OKID) [13], [18] assumes an observer (1)
Y0 = Y0 = 0, Y1 = Y1 , (20a)
structure of the treated system. The system is excited by a
(1)

i1
(2)
known (measured) excitation signal u(k) such that it reach the Yi = Yi Yj Yij , i = 2, . . . , p, (20b)
system states. The observer structure allows compressing the j=1
data and this improves system identication results in practice
p
(2)
(pulse response decay much faster than that of the original Yi = Yj Yij , i = p + 1, . . . , l 1. (20c)
one). The system MP are determined recursively from the j=1
observer MP. In the following, we will summarize the OKID Note that data are taken from a different time step p and, for
technique in brief. practical purposes, the identication process can be done by
Consider the following observer structure, where an ob- taking an arbitrary time-window of data from the actual system
server term has been added to the right-hand side of (1): under operation. Once the system MP determined, ERA is used
x(k + 1) = A x(k) + B v(k) to obtain a system realization as follows.
(11)
y(k) = C x(k), B. Computation using the ERA
[ ]
with A = A + GC and B = B G . The variable v(k) A Hankel matrix [11] is formed using the MP from (20),
Rr+m is employed in order to take into account the measures as follows:

from system and observer: Yi Yi+1 Yi+s1
[ ]T Yi+1 Yi+2 Yi+s
v(k) = uT (k) y T (k) . (12)
H(i 1) = . . . .. , (21)
.. .. .. .
Note that u(k) is the input-signal known or measured used as
excitation of the system. The observer is supposed to have a Yi+q1 Yi+q Yi+q+s2
dead-beat structure, so A is asymptotically stable for p samples where s, q > n are positive integers chosen arbitrary, therefore
(Ap 0), where k p and it represents the number of H(i 1) Rmqrs . This matrix is, by singular value decom-
independent MP. The linear regression is position (SVD), related to A by the following relationship:
= Y V , (13) H(i 1) = P Ai1 Q, (22)
with dimensions of each component as follows: mlp
, where the observability (P ) and the controllability (Q) of the
Y m[(m+r)p+r] , V [(m+r)p+r]lp , where l is the number system are
of data samples, and
[ ] C
= y(p) y(p + 1) y(l 1) , (14) CA [ ]
[ ]
Y = 0 C B C AB C A B , p1 P = . , and Q = B AB As1 B . (23)

(15) . .
u(p) u(p + 1) u(l 1) CAq1
v(p 1) v(p) v(l 2)
Then, for i = 1 in (22) and using SVD, the following
v(p 2) v(p 1) v(l 3)
V = . (16) relationship is obtained:
.. .. .. .. [ ]
. . . .
T Sn 0
v(0) v(1) v(l p 1) H(0) = U SV , S =
0 0
, (24)

127
where the subindex n represents the order of the system
(known a priori ). The stronger values of Sn are taken as
the more excited controllable (also observable) system com-
ponents, representing the system dynamics. Therefore,
H(0) = Un Sn VnT , (25)
where Un and Vn are constituted by the rst n columns of
matrices U and V of (24), respectively. The triplet (A, B, C)
is obtained using (22) for i = 2, and (25), as follows:
A =Sn1/2 UnT H(1)Vn Sn1/2 , (26a)
B =Sn1/2 VnT Er , (26b)
T
C =Em Un Sn1/2 , (26c)
where selection matrices Em and Er are obtained from (22)
and (25), and dened as Fig. 1: Functional blocks for data-driven recongurability-
[ ]
Em = Im 0m 0m , Em T
Rmmq , (27a) based index computation.
[ ]
Er = Ir 0r 0r , Er R
T rrs
. (27b)
C. Evaluating indexes For illustrative purposes, the control law applied to this system
Since the nominal system is supposed to be known (model- is based on a linear quadratic regulator, where weighting ma-
based) [3], the dynamics is known and hence the indexes be trices values, Q = 0.01I3 , R = 0.1I2 , give more importance
compared for supervision purposes. This is illustrated by the to the control cost rather than regulating the system states and
blocks of the Fig. 1. From the input/output data the observer then avoiding great control effort [22]. The state feedback gain
MP are computed and then, using OKID technique, the system is: [ ]
MP. Afterwards the identied representation (A, B, C) is 2.3882 4.6741 0.0685
K=
obtained through ERA. From this representation, the Han- 0.2151 34.9663 0.0763
kel singular values or a balanced identied model (triplet
Abe , Beb , Ceb in Fig. 1) are computed (see Section III), from and, in order to track a given speed reference signal, the
which the recongurability-based index is nally computed. feedforward gain matrix is:
The values obtained on-line each required time after a fault has [ ]
1.1202 0.5168
occurred, are then compared to the analytical ones, or used in N= .
10.0013 0.2105
the supervision stage in order to evaluate the recongurability
in the actual system operation conditions. In order to evaluate the recongurability, for actuator faults
The following example is intended to show the usefulness matrix B is viewed as B = [b1 b2 ], whereas for sensor
of the recongurability-based index evaluation in off-line and faults matrix C is viewed as C = [c1 c2 c3 ]T , and each
on-line cases. of this vectors is affected by loss of effectiveness. Therefore
V. A PPLICATION E XAMPLE the index Q% is evaluated w.r.t. sensor and actuator faults.
The regarded faults are sudden changes in the effectiveness
A DC motor in separately-excited topology [19] is used for
factor at a unknown occurrence time. An ideal FDI module
academic purposes in order to illustrate the ideas previously
provides the fault occurrence time and res the signal starting
presented. The values proposed in [20] are considered, as well
the algorithm for recongurability evaluation.
as the results from [21] for comparison purposes related to
Firstly, the off-line analysis of sensor faults is considered.
actions to take in the event of fault occurrence. The inputs to
After obtention of the balanced form (5) using the matrix
the model are armature (rotor) voltage u1 , eld (stator) voltage
Tb , the controllability gramian is calculated and then used to
u2 . The states (viewed at the output) represent armature current
evaluate % using (7) for different values of i . The nominal
x1 = ia , eld current x2 = if , and angular velocity x3 = .
value is %n = 0.2406. Using (9) the index Q% is computed
For the operating point around y0 = [5 0.1 100]T , u0 =
for each set of sensors, i.e., {ia , }, {if , }, {ia , if }.
[53.7388 47.3795]T and h = 0.01 s, the linearized model is
Fig. 2 shows the plot of index Q% for the rst one. Axes x, y
represented by the following matrices:
show the variation for each , whereas z axis represents the
0.9791 1.8411 0.0019 Q% values. As it can be expected, the value goes to zero when
A = 0.0 0.9208 0.0 , both sensors lose their affectiveness. This plot is similar to that
0.8790 40.6421 0.9875 of set {if , }, not showed here by space limitations. Note
(28)
0.0054 0.0002 that system tolerates faults on the armature current sensor ia
B = 0.0 0.0002 , C = I3 . better than speed sensor , then the system is more sensitive
0.0024 0.0037 to faults occurring in this last sensor.

128
(a) Q% (b) Limited Q% (Qad = 40%)

Fig. 2: Recongurability evaluation for sensor faults: set {ia , }.

Consider limiting the operation of the system to a given q, s to form the Hankel matrix in (21) are chosen to be q = 8
value of Q% under a given faulty situation. For this control and s = 2q. The upper plots of Fig. 4 show the output signals
system, an admissible value Qad = 40% is settled in order (ia , if , , in the a), b), c) plots). As noted, these signals are
to establish fault accommodation, control reconguration or used only at times required to perform the computations of
stopping operation. Values over 40% will be accepted, and Q% , in this case in an interval where there is no fault and then
values below 40% will be rejected, as shows Fig. 2b. With when the FDI module has detected the occurrence of the fault
this admissible value is clear from the plot that speed sensor at t = 8.0 s. After that, the identication algorithm is invoked.
is more penalized. Furthermore the results obtained in [21] The output signals in this faulty case (dotted lines in Fig. 4
show that loss of effectiveness in both sensors up to 50% is a), b), c)) are compared with the nominal ones.
acceptable for fault accommodation purposes. However, note The last plot (Fig. 4 d)) shows the evolution of the index
from Fig. 3a and with admissible value Qad = 40%, that for
faults occurring in both current sensors ia , if , accommodation
of faults could not be the better choice for the loss of Fault occurrence
effectiveness considered previously. In this case sensor ia is Time window Time window

more penalized than sensor if . normal


faulty
Consider now actuator faults. Similar to sensor faults, the
a)
index Q% for different values of i is computed, from which
a

the plot of rejected/accepted values for Qad = 40% is obtained


and shown in Fig. 3b. Faults on actuator u2 impact the index
in major degree than actuator u1 . This is explained by the fact normal
faulty
that the excitation of the motor depends on the former [19].
b)
The on-line (no calculation at each time step using persistent
excitations is required) evaluation of Q% is done using ERA
with observer in the fault free and faulty cases. Whilst the
ERA may be invoked multiple times during simulation, a time normal
faulty
window is required to get data and then to calculate % using reference
c)
(7) for the identied triplet (A, B, C). Again the balanced
transformation using matrix Tb is required to set the previous
triplet under the form (4). Computation is done two times:
rst in nominal case (at time 5.0 s) and the second one when
a loss of effectiveness of 50% has occurred (at time 8.0 s)
in the second actuator u2 (in the eld circuit). The reference 0.5451
Q

Qad=0.4 d)
speed signal to track is the trapezoidal one showed in Fig. 4
c) (dashed line).
The ERA computations are done with 0.5 s taken as the
data length of the time window, as can be viewed in Fig. 4.
Also, the considered number of MP is 10, and the integers Fig. 4: Outputs (ia , if , ), and Q% evolution.

129
(a) Evaluation of the set {ia , if } (b) Evaluation of the set {u1 , u2 }

Fig. 3: Recongurability-based index evaluation: Limited Q% (Qad = 40%).

Q% in order to verify the change when the fault has occurred. [6] B. Moore, Principal component analysis in linear systems: Control-
The value in this faulty condition is Q% = 0.5451, which lability, observability, and model reduction, IEEE Transactions on
Automatic Control, vol. 26, no. 1, pp. 1732, Feb. 1981.
corresponds to that obtained off-line. Taking into account the [7] N. E. Wu, K. Zhou, and G. Salomon, Control recongurability of linear
Fig. 3b this value is admissible in order to continue with time-invariant systems, Automatica, vol. 36, no. 11, pp. 17671771,
the system operation by using an accommodation strategy Nov. 2000.
[8] M. Staroswiecki, On recongurability with respect to actuator failures,
allowing to increase the control signal to compensate the loss in Proc. of the 15th Triennial World Congress of the IFAC 2002,
of actuation, as it was done in [21]. Barcelona, Spain, July 2002, pp. 775780.
[9] Z. Yang, Recongurability analysis for a class of linear hybrid systems,
VI. C ONCLUSION in Proc. of the 6th IFAC SafeProcess06, Beijing, China, Aug. 30-Sept.
1 2006, pp. 10331038.
In this paper, evaluation of the system control recong- [10] F. Chowdhury and W. Chen, Fault monitoring in the presence of fault-
urability from input/output data (data-driven) and an index tolerant control, in Proc. of the 6th IFAC SafeProcess06, Beijing,
China, Aug. 30-Sept. 1 2006, pp. 13211326.
based on the control recongurability have been proposed. [11] K. Zhou, J. Doyle, and K. Glover, Robust and optimal control. Engle-
Control recongurability is a property related directly to the wood Cliffs: Prentice-Hall, 1996.
controllability gramian and the controllability of the system. [12] J.-N. Juang and R. Pappa, An eigensystem realization algorithm for
model parameter identication and model reduction, AIAA Journal of
This relation characterizes those states reachable by a specied Guidance, Control, and Dyn., vol. 8, no. 5, pp. 620627, Sep.-Oct. 1985.
energy input. Founded on the same base, this paper proposed [13] J.-N. Juang, M. Phan, L. Horta, and R. Longman, Identication of
to take advantage from the ERA/OKID technique to obtain a observer/kalman lter markov parameters: theory and experiments,
AIAA Journal of Guidance, Control, and Dynamics, vol. 16, no. 2, pp.
representation allowing on-line computation of the index based 320329, March-April 1993.
on recongurability, and further comparaison with the nominal [14] Y. Zhang and J. Jiang, Active fault-tolerant control system against
one obtained off-line. partial actuator failures, IEE Proceedings of Control Theory and
Applications, vol. 149, no. 1, pp. 95104, Jan. 2002.
A DC motor simulation example has been presented in [15] M. Mahmoud, J. Jiang, and Y. Zhang, Active Fault Tolerant Control
order to show the ideas proposed and the pertinence using Systems, ser. Lecture Notes in Control and Information Sciences. UK:
the index for supervision of fault tolerance tasks. The index Springer, 2003, vol. 287.
[16] S. X. Ding, Model-based Fault Diagnosis Techniques. Berlin: Springer
might be employed not only in the design phase of a fault Berlin Heidelberg, 2008.
tolerant control system, but also after its application in order [17] M. Verhaegen and V. Verdult, Filtering and System Identication.
to evaluate its tolerance to faults. Cambridge: Cambridge University Press, 2007.
[18] J.-N. Juang, Applied system identication. Englewood Cliffs: Prentice
Hall, 1994.
R EFERENCES [19] J. Chiasson, Modeling and high performance control of electric ma-
[1] Y. Zhang and J. Jiang, Bibliographical review on recongurable fault- chines. UK: Wiley-IEEE Press, 2005.
tolerant control systems, Annual Reviews in Control, vol. 32, no. 2, pp. [20] D. Campos-Delgado, E. Palacios, and D. Espinoza-Trejo, Passivity
229252, Dec. 2008. based control of nonlinear dc motors congurations and sensorless
[2] J. Jiang, Fault-tolerant control systems. An introductory overview, Acta applications, in Proc. of the International Symposium on Industrial
Automatica Sinica, vol. 31, no. 1, pp. 160174, Jan. 2005. Electronics, ISIE07, Vigo, Spain, June. 4-7 2007, pp. 33793384.
[3] R. Isermann, Fault-diagnosis systems: An introduction from fault detec- [21] D. Campos-Delgado and D. Espinoza-Trejo, Active fault tolerant
tion to fault tolerance. Berlin: Springer, 2006. scheme for variable speed drives under actuator and sensor faults, in
[4] M. Blanke, M. Kinnaert, J. Lunze, and M. Staroswiecki, Diagnosis and Proc. of the International Conference on Control Applications, CCA08,
Fault-Tolerant Control, ser. Control Systems Series. Springer, 2006. San Antonio, Texas, USA, Sep. 3-5 2008, pp. 474479.
[5] C. Frei, F. Kraus, and M. Blanke, Recoverability viewed as a system [22] B. Anderson and J. Moore, Optimal control: Linear quadratic methods.
property, in Proc. of the European Control Conference, ECC99, USA: Prentice Hall, 1990.
Karlsruhe, Germany, June 1999, pp. CDROM.

130

S-ar putea să vă placă și