Вы находитесь на странице: 1из 8

4th World Conference on Structural Control and Monitoring 4WCSCM-066

ANALYSIS OF A DISCRETE FIRST-ORDER MODEL REFERENCE ADAPTIVE


CONTROLLER DISCRETIZED BY THE ZERO-ORDER-HOLD DISCRETE EQUIVALENT
O.S. Bursi, N. Tondini and L. Vulcan
University of Trento, Trento, 38050, ITALY
oreste.bursi@ing.unitn.it, nicola.tondini@ing.unitn.it, leonardo.vulcan@ing.unitn.it

D.P. Stoten
University of Bristol, Bristol, BS8 1TR, UK
d.p.stoten@bristol.ac.uk

Abstract

Structural active control has ever greater interest, as it allows increasing the performance and the life of a structure in the face of
environmental actions. In detail, the adaptive control involves techniques that permit to control systems whose characteristics are
uncertain or can vary with time. Nowadays, control systems are digitally implemented, then the control algorithm must be written
in a discrete form. The research made has the aim to analyse a model reference adaptive controller (MRAC) discretized by the
zero-order-hold (ZOH) discrete equivalent. In an MRAC the characteristics of the system are given in terms of a linear reference
model and the goal of the controller is to match the response of the plant with that of the reference model by means of an
adaptive mechanism conceived to compensate for system nonlinearities. Among the MRAC controllers, the Minimal Control
Synthesis (MCS) algorithm is analysed and applied for instance to a linear first-order SDoF system. The single step/stage ZOH
method is used, as it is suitable to be implemented for real-time applications. Both stability and accuracy analyses are carried out
in order to identify the performance of the MCS algorithm. The analysis is conducted by system linearization at steady-state
conditions through a physical insight approach. The unavoidable presence of a delay in the computation of the control law leads
to consider its effect by carrying out the comparison of two cases: (i) no delay; (ii) delay equal to one time step. Finally,
numerical findings are validated by means of simulations.

Introduction

The use of adaptive control techniques has become a topic of increasing interest in recent years as
adaptive control can be used to control plants whose parameters are unknown or uncertain (strom et al.,
1995). In the context of control theory some recent books and papers relevant to direct methods can be
found (Sastry, 1999; Wagg, 2003), while for structural control several applications have been made
(Housner et al., 1997). When a direct reference model controller is used, it can be also applied to systems
where the details of the plant cannot be fully known a priori or are time varying. Using this type of
algorithms assuming zero initial conditions for the controller gains, i.e. without the need of the plant
parameter knowledge, has become known as the Minimal Control Synthesis (MCS) approach (Stoten and
Benchoubane, 1990). Basing adaptive control schemes on a reference model enables the controlled
system to behave like the model itself. This type of approach has been applied to a wide range of plants
including nonlinear systems (Wagg, 2003). As the approach is primarily based on linear control theory,
being the reference model usually a linear one, the effect of nonlinearities and/or disturbances in
nonlinear systems is compensated for by the adaptive nature of the controller. Nowadays, controllers are
implemented on digital computers, and therefore, control systems must be directly written in digital form
and/or transformed from continuous time to discrete time; moreover, if possible, they must run in real-
time and compensate for the influence of time delay (Chu et al., 2002). Particular attention will be given
to the time delay issue because it sensitively affects the control system behaviour and stability. In general,
the total time delay in a control system can be divided in two contributions. The first part is referred to as
fixed time delay due to the on-line data acquisition, filtering, manipulation of digital data inside the
digital control processor, calculation of required control force and signal transmission from the computer
to the actuator. The second part usually depends on the particular dynamics of the actuators interacting
with the controlled system, which is referred to as floating delay time that can be adjusted by explicitly
combining the actuator dynamics with that of the structure (Chu et al., 2002). Besides, time delays are

Bursi, Stoten, Tondini and Vulcan 1


often the source of instability of the system (Xu and Lam, 2005), so this paper will investigate the
behaviour of the system by considering one time step delay in the discrete control law. The selection of
the best sample rate for a digital control system still represents a problem. Thus generally, the slowest
sample rate that meets all performance specifications is sought. The paper is organised as follows.
Initially, the MCS algorithm formulation and the adaptive mechanism for nonlinearity compensation will
be presented. Then, the MCS algorithm with one time step delay in the control law, discretized by
implementing the ZOH integrator, will be investigated by means of stability and accuracy analyses.
Afterwards, simulations capable to show the effect of the delay on the performance of the control system
will be presented. Finally, the main conclusions along with comments on future perspectives will be
drawn.

The Minimal Control Synthesis algorithm

This section introduces the features of the adaptive MCS controller exploited in this paper; this controller
is characterized by an adaptive portion, conceived to compensate for system nonlinearities. A generic
nonlinear plant can be described by the following equations:
x (t ) = x(t ) + Bu(t ) + f NL (x(t ), t )
(1)
y (t ) = Cx(t )
where A is the nn plant matrix, B is the nm input matrix, x is the n-dimensional state vector, u is the m-
dimensional control input vector and fNL represents the nonlinear and time varying terms. y and C are the
output vector and plant output matrix, respectively, each having appropriate dimensions. A linear system
with the same number of states, inputs and outputs is selected as reference model. The goal of the
controller is to match the characteristics of the actual nonlinear system with those of the reference model.
The reference model has the following form:
x M (t ) = M x M (t ) + B M r (t )
(2)
y M (t ) = Cx M (t )
where AM and BM are nn and nm constant matrices, respectively, xM is the reference model state vector,
r is the m-dimensional reference input vector, and yM is the reference model output. The error vector, xE,
can be defined as the difference between reference model, xM, and plant state, x. The desired result for the
controller is to drive the steady-state error such that limt+xE=0, when this limit exists. If the plant is
time varying or nonlinear, then the fNL vector influences the controller. For this reason, the Model
Reference Adaptive System (MRAS) theory introduces a time varying controller:
u (t ) = - (K - K (t )) x (t ) + (K R + K R (t )) r (t ) (3)

where K and KR are time varying gain adjustments. Substituting the control law (3) in (1) and after
some manipulations, we get
x E ( t ) = A M x E ( t ) + ( A M - A + BK ) x ( t ) +
(4)
(
+ ( B M - BK R ) r ( t ) - B ( K R ( t ) r ( t ) + K ( t ) x ( t ) ) + f NL ( x ( t ) , t ) )
K and KR are selected so that the second and third terms in (4) are cancelled out. The error equation now
reduces to:
x E ( t ) = A M x E ( t ) - B ( K R ( t ) r ( t ) + K ( t ) x ( t ) ) - f NL ( x ( t ) , t ) (5)

Bursi, Stoten, Tondini and Vulcan 2


The values of K and KR must be selected to cancel out the nonlinear terms, but they cannot explicitly
solved because x is unknown. A stability proof for these controllers has been developed by using the
hyperstability theory (Landau, 1979), which deals with the stability of systems that can be broken into a
linear feed-forward loop, which meets the strictly positive real condition, and a nonlinear feedback loop
that satisfies the Popov criterion for hyperstability. This problem is solved by the proper selection of the
output error matrix CE, used to define the output error yE(t) = CE xE(t), according to the Lyapunov
problem. A solution that satisfies the hyperstability condition yields expressions for K and KR (Landau,
1979):
t
K ( t ) = y E ( ) xT ( ) d + y E ( t ) xT ( t )
0
t
(6)
K R ( t ) = y E ( ) r T ( ) d + y E ( t ) rT ( t )
0

where and are two positive constants, that must be selected before the solution of the above equations.
In order to ensure the stability of the closed loop system, AM must have eigenvalues in the left hand side
of the complex plane and should also have a Luenberger-type controllable canonical structure to ensure
that Erzbergers conditions are satisfied (Landau, 1979). The adaptive controller presented above requires
the knowledge of the plant characteristics, A and B, which allows evaluating the constant values of K and
KR. This drawback can be eliminated by using the MCS algorithm which derives from the MRAS
formulation, but it assumes zero initial gains K and KR and, consequently, unknown plant parameters
(Stoten and Benchoubane, 1990). Figure 1 shows the block diagram of the MCS controller with its
essential components: the adaptive block; the input signal generation, u; the plant; together with a parallel
reference model; and a common reference signal, r. The only elements that are needed to know are: the
reference model parameters AM and BM; the structure of the plant with the degree of freedom and order
and the sign of the coefficients of B, that usually is assumed positive. Therefore, the control law becomes

u (t ) = K (t ) x (t ) + K R (t ) r (t ) (7)

For the value of the output matrix CE, Stoten and Neild (2003) proposed a pragmatic solution for first-
order (CE=[4/ts]) and second-order (CE=[4/ts 1]) one Degree-of-Freedom systems (DoFs), where ts is the
time settling of the reference model. This solution entails an exact pole-zero cancellation.

xm ym
xm=Amxm+Bmr C
+ xe ye
Ce
u=Kx+K Rr
r + y
x=Ax+Bu x
C
+

Figure 1. Block diagram of the classical Minimal Control Synthesis algorithm.

Discrete formulation by means of the ZOH method

Since usually the controller is implemented in a digital fashion, the continuous controller developed in the
previous section must be discretized. This transformation is performed by means of the ZOH sampling
process which exhibits several features among which, exact solution at sampling points for linear time

Bursi, Stoten, Tondini and Vulcan 3


invariant (LTI) systems and real-time compatibility. The discrete-time form of the reference model is
defined by the mapping
x M [ k + 1] = M x M [ k ] + M r [ k ] (8)

resulting in one-step one-stage method, where


t
M = e tA M
and M = e A M d B M (9)
0

The sampling period or time interval is t= tk+1 - tk where the discrete time variable is the integer k. The
discrete control law equation for the MCS controller reads:

u [k ] = K [k ] x [k ] + K R [k ]r [k ] (10)

with the adaptive gains,


K [ k ] = K [ k 1] + y E [ k ] xT [ k ] y E [ k 1] xT [ k 1]
(11)
K R [ k ] = K R [ k 1] + y E [ k ] rT [ k ] y E [ k 1] rT [ k 1]
where =t. For the assumptions of the MCS algorithm, the initial conditions read K[1]=0 and
KR[1]=0. In order to complete the characterization of the MCS controller, the eigenvalues of the
reference model, and from the adaptation equations, just as CE must be selected. The reference model
must be stable, i.e. the eigenvalue moduli must be less than one, and should be selected so as not to
exceed the system capability. The values of and are arbitrary positive numbers and are selected by
trial and error. However, an increase of means a reduction of the settling time of the adaptation, while a
reduction of the / ratio improves the damping. The ratio / can be set to 0.1 as compromise between
the speed of adaptation mechanism and stability limit (Vulcan, 2006). The final values of and are not
critical, but they cannot be increased indefinitely because that may magnify noise within the loop. The
selection of the error vector weighting matrix CE is also relatively arbitrary and the selected values have
been already introduced in the previous section for the continuous time system. Due to the presence of the
adaptive process, the MCS algorithm defines a nonlinear dynamical system; therefore, for simplicity, we
will consider first order linear systems only both for the plant and the reference model, and we will
perform the analysis on SDoF model equations. As a result:
x [ k + 1] = A ' x [ k ] + B ' u [ k ]
(12)
xM [ k + 1] = AM' xM [ k ] + BM' r [ k ]

where: A' = e t / T , B ' = 1 e t / T , AM' = et / TM , and BM' = 1 et / TM owing to the ZOH sampling; T and
TM are the plant and the reference model time constants, respectively; the low frequency gain for the plant
and the model are assumed equal to 1; and TM is related to the approximate settling time ts by TM=ts/4.
Furthermore, the control signal and the reference input, that is assumed as a step function, read,
u [k ] = K [k ] x [k ] + KR [k ] r [k ] (13)
r [ k ] = 0 for k < 0, r [ k ] = 1 for k 0 (14)

respectively. In (13), we have implicitly assumed that the processing time is negligible with respect to
the time step t and, therefore, the control signal u, generated by the MCS algorithm, is obtained
instantaneously and held constant over t. Furthermore, (11) are written in scalar form and CE= 4/ts is
selected according to the proposed solution. As just stated, the control variable u[k] and the state variable

Bursi, Stoten, Tondini and Vulcan 4


x[k] are assumed to be computed at the same time. Evidently, this is an approximation as the time delay
during the evaluation of the control signal is unavoidable. In this sense the behaviour of the system will
be analysed by considering one time step delay in the discrete control law and, consequently, the
comparison between the case with and without delay will be presented. In these conditions, the control
law of the MCS controller reads,
u[k ] = K [k 1]x[k 1] + K R [k 1]r[k 1] (15)

We emphasize that (15) entails another approximation, since in reality the delay may be a fraction of t;
however it is always possible to manage the delay to be equal to the time step, such that (15) results to be
correct. Therefore, the analyses provided hereafter allow the effect of a delay in the control law
computation on the system to be understood.

Stability and accuracy analysis

Following Vulcan (2006) the discrete-time equations that rule the system with delay can be arranged as:
xE [ k +1] xE [ k ] B 'm r[ k ] '
AM 0 AM' A ' B ' x[ k 1] B ' r[ k 1]

xE [ k ] xE [ k 1] 0 1 0 0 0 0

x[ k +1] = C MD x[ k ] + 0 , C MD = 0 0 A' B ' x[ k 1] B ' r[ k 1] (16)
K k 1
K[k ] [ ] 0 x[ k ]C x[ k 1]CE 0 1 0
E

K R [ k ] K R [ k 1] 0 r[ k ]CE r[ k 1]CE 0 0 1

In order to reduce the complexity of the analysis, it will be locally developed by linearizing the dynamical
system by a physical insight approach around an operating point. The linearization by physical insight is
simply carried out by substituting the values of the operating point in the amplification matrix, which in
turn is arranged in a way to reduce the complexity of the analysis. One can note that the system
linearization could be obtained through a more rigorous linearization procedure based on a Taylor series
expansion about the operating point (Dutton et al., 1997), but this approach is more burdensome than the
physical insight because it requires preliminary simulations in order to estimate K[k] at steady state.
Considering a unit step input, the operating point at steady-state conditions reads: xE[k-1]=xE[k]=0, x[k-
1]=x[k]=1, r[k-1]=r[k]=1. In this way, substituting the operating point in the amplification matrix (16), we
obtain the amplification matrix at steady state.
AM' 0 AM' A ' B ' B '

1 0 0 0 0
(17)
CMD , ss = 0 0 A' B' B'

CE CE 0 1 0
C CE 0 0 1
E
Firstly, CMD,ss is assembled in order to avoid having terms depending on xE, as they would become null
once substituted the operating point in CMD,ss. Secondly, the operating point contains the steady-state
values of K and KR that are unknown, so every term exhibiting the values of K and KR is moved to the
augmented vector at step k. These choices allow an amplification matrix at steady-state to be obtained
easy to be analysed by introducing the following non-dimensional variables: t/TM, /(4TM), T/TM and
/(4TM). It can be proved that the stability of the system is assured when the modulus of the eigenvalues
is less than one (Bursi et al., 2006). Considering that it is possible to determine the Sampling interval -
Gain (SG) space, which represents the asymptotic stability regions by varying t/TM and /(4TM), for
/(4TM)=0.1 and r[k]=1, as it is depicted in Figure 2. These spaces may also be interpreted as operational
tools that allow the optimal combination of t-/ values to be selected; conversely they are usually set
by trial and errors, even if the T/TM ratio is unknown. The advantage of these spaces is that they permit

Bursi, Stoten, Tondini and Vulcan 5


the operators to determine a priori information on the behaviour of the sampled-data control system.
From Figure 2 it is evident that, if the control law u[k] is computed with one time step delay, the stability
limit reduces. Such a result was expected as the MCS controller is not able to promptly modify the system
parameters due to the delay in the evaluation of the control law. These findings underline the need of
reducing the time of the control law evaluation or to assume small ratios t/TM, either by using small t
or large TM, to increase the stability of the system for delay effects. The results found on system stability
are confirmed by the plot of the SG space shown in Figure 2. Moreover, it highlights a greater reduction
of the stability regions when the ratio T/TM increases. An increase of the ratio T/TM means that the plant
becomes ever slower, hence, a loss of time in the control law computation, due to the delay, is more
burdensome in terms of stability. It may be also observed that the complex limit of the system without
delay, which is independent from the plant characteristics, disappears; hence, the stability boundaries
depend on the plant.
100 100
/(4T ) /(4Tm)
m

10

10
1

0.1
1

0.01

0.1 0.001
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0.001 0.01 0.1 1
T/Tm=1 delay T/Tm=10 delay T/Tm=1 t/Tm T/Tm=1 delay T/Tm=10 delay T/Tm=1 /(4Tm)
T/Tm=10 complex conjugate T/Tm=10 complex conjugate

Figure 2. The SG space for the MCS with one time Figure 3. The GS domain for the MCS with one
step delay, r[k]=1 and /(4TM)=0.1. time step delay, r[k]=1 and t/TM=0.1

An attentive reader may observe from Figure 2 that the behaviour of the system with delay starts to be
quite independent from the plant characteristics when t/TM approaches 0.4 and from this value onwards.
Notice that 0.4 corresponds to the vertical asymptote for the system without delay. Similar conclusions
can be drawn from the Gain Space (GS) domain for variable /(4TM) and t/TM=0.1 depicted in Figure
3, where the system behaves independently from the ratio T/TM as long as /(4TM) 0.025; this value
corresponds to the vertical asymptote of the complex limit for the system without delay. It is worthwhile
presenting also the algorithmic damping ratio and the algorithmic damped numerical frequency , as
defined by Bursi et al. (2006), in order to show how the MCS algorithm influences the system
performance. It may be observed from Figures 4 and 5 that both and introduced by the controller
decrease for the same parameter combination of Figure 2.

= 1 delay = 1
2.5 4 Tm 4 Tm

= 1 delay

= 1
4 Tm 4 Tm
= 10 delay = 10
4 Tm 4 Tm

2 = 10 delay = 10
4 Tm 4 Tm 3
4

1.5

2
1


0.5 4

t t
Tm Tm
0.2 0.4 0.6 0.8 1 0.2 0.4 0.6 0.8 1

Figure 4. as function of t/TM for the MCS Figure 5. as function of t/TM for the MCS
algorithm with one time step delay, T/TM=10 and algorithm with one time step delay, T/TM=10 and
r[k]=1. r[k]=1.

Bursi, Stoten, Tondini and Vulcan 6


Performance of the MCS algorithm with one time step delay

In order to study the performance of the MCS algorithm with one time step delay, the mean square
tracking error XE[k] and the tracking error bound XE,sup[k] are used. They are suggested and defined in
Bursi et al. (2006). From Figure 6, one can note that XE increases faster for the MCS with delay when
t/TM becomes large; this is particularly evident when t/Tm approaches 0.4. This behaviour is due to the
fact that the system with delay at that value is getting unstable, see Figure 2, and, therefore, it is much
closer to the stability limit than the standard MCS. This trend can be also clearly observed in Figure 7 for
T/Tm=10, this figure reports the tracking error bound XE,sup as function of t/Tm.

xE 0.01 0.1 1 xE,sup 0.01 0.1 1


1.E+00 1.E+01

1.E-01
1.E+00

1.E-02
1.E-01
1.E-03
1.E-02
1.E-04
1.E-03
1.E-05

1.E-04
1.E-06

1.E-07 1.E-05

1.E-08 1.E-06
t/Tm t/Tm

Figure 6. Convergence of XE at t=ts for the MCS Figure 7. Convergence of XE,sup at t=2ts for the
with one time step delay, T/TM=10 and r=1: the full MCS with one time step delay, r=1 and T/TM=10:
line correspond to /(4TM)=0.1; the dashed to the full line corresponds to /(4TM)=0.1; the
/(4TM)=1; the dotted to /(4TM)=10. dashed to /(4TM)=1; the dotted to /(4TM)=10.

Representative numerical simulations

The simulations presented hereafter intend to validate some of the analytical findings. It is demonstrated
that the MCS with one time step delay determines a reduction of the stability limits as well as the
algorithmic damping introduced by the controller (Vulcan, 2006). For instance, the evolution of the MCS
controlled system with and without delay is depicted in Figures 8 and 9, respectively, assuming
/(4TM)=10 and t/TM=0.1. They highlight that the system controlled by the MCS with delay starts to
become unstable, while the system controlled by the standard MCS keeps staying stable. This could be
verified from the SG space of Figure 2. Moreover, Figure 8 shows how the system is less damped out and
the oscillations become larger, due to a reduced value of the algorithmic damping depicted in Figure 4.
Conversely, the system without delay depicted in Figure 9 does not exhibit any particular change in this
respect.

Conclusions and perspectives

This paper has dealt with the convergence and the steady-state analysis of a discrete first-order MCS
controller sampled with the one-step one-stage ZOH discrete equivalent. In detail, it was assumed that the
plant to be controlled was a first-order system and the communication between the plant and the
controller was zero-order sampled. In the proposed analysis, the control signal was generated considering
the effect of a time delay, at this stage equal to one time step t, as the delay is always present in physical
devices. The analysis was performed for the case where the demand corresponds to a unit step input. Due
to the system nonlinearity, the system was linearized by means of a novel physical insight approach. This
analysis allows two gain space domains to be determined in order to define the region of local stability.
Moreover, the accuracy analysis has provided insight into the range of adaptive control weights that result

Bursi, Stoten, Tondini and Vulcan 7


in optimal performance of the MCS controller. This analysis also highlighted a possible approach to a
priori selection of the time step and adaptive weighting values.

14 14

12 12

10 10

8 8

6 6

4 4

2 2

0 0
0.0 0.5 1.0 1.5 2.0 2.5 0.0 0.5 1.0 1.5 2.0 2.5
-2 -2

-4 -4

-6 -6
x xe K Kr Time x xe K Kr Time

Figure 8. Evolution of the MCS controlled system Figure 9. Evolution of the MCS controlled system
with delay for r[k]=1, T/TM=10, /(4TM)=10, for r[k]=1, T/TM=10, /(4TM)=10, /(4TM)=0.1, and
/(4TM)=0.1, and t/TM=0.1. t/TM=0.1.

In terms of future perspectives, the analysis proposed in this paper and performed for a first-order system,
could be developed for second-order systems, which are typical present in structural systems, in order to
complete the convergence and the steady-state analysis of the discrete MCS controller. The effect of the
delay reduces the stability limit and the effectiveness of the numerical damping; this effect represents an
important characteristic of discrete control system that can be further deepened in control problems
considering the effect of variable delay.

Acknowledgements

The authors thank the Department of Mechanical and Structural Engineering for the SMART STRUCTURES project, that
sponsored the research project by means of grants. However, opinions expressed in this paper are those of the writers, and do not
necessarily reflect those of the sponsor.

References
strm, K.J. and B. Wittenmark (1995), Adaptive Control, Addison-Wesley Publishing Company, New York.
Bursi O.S., D.P. Stoten and L. Vulcan (2006), Convergence and frequency-domain analysis of a discrete first-order model
reference adaptive controller, Structural Control and Health Monitoring, (in print).
Chu S.Y., T.T. Soong, C.C. Lin, and Y.Z. Chen, (2002) Time-delay Effect and Compensation on Direct Output Feedback
Controlled Mass Damper Systems, Earthquake Engineering and Structural Dynamics, 31:121-137.
Dutton K., S. Thompson and B. Barraclough (1997), The Art of Control Engineering, Addison Wesley Longman, Harlow.
Housner G.W., L.A. Bergman, T.K. Caughey, A.G. Chassiakos, R.O. Claus, S.F. Masri, R.E. Skelton, T.T. Soong, B.F. Spencer
BF, Jr., and T.P. Yao (1997), Structural control: past, present, and future, J. of Engineering Mechanics, 123(9): 897-971.
Landau Y.D. (1979), Adaptive Control The Model Reference Approach, Marcel Dekker Inc., New York.
Sastry S. (1999), Nonlinear Systems. Analysis, Stability and Control, Springer Verlag, Berlin.
Stoten D.P. and H. Benchoubane (1990), Empirical studies of an MRAC algorithm with minimal controller synthesis,
International Journal of Control, 51(4): 823-849.
Stoten D.P. and S.A. Neild (2003), The error-based minimal control synthesis algorithm with integral action, Proceeding
Institutions of Mechanical Engineers- Part I, 217: 187-201.
Vulcan L. (2006), Discrete-time analysis of integrator algorithms applied to S.I.S.O. adaptive controllers with Minimal Control
Synthesis, Ph.D. thesis, Department of Mechanical and Structural Engineering, University of Trento, Trento.
Wagg D.J. (2003), Adaptive control of nonlinear dynamical systems using a model reference approach, Meccanica; 38(2): 227-
238.
Xu S. and J. Lam (2005), Improved Delay-Dependent Stability Criteria for Time-Delay Systems, IEEE Transactions on
Automatic Control, 50(3): 384-387.

Bursi, Stoten, Tondini and Vulcan 8

Вам также может понравиться