International Conference on Control, Automation and Systems 2008

Oct. 14-17, 2008 in COEX, Seoul, Korea

Simple Model Matching Control of Nonlinear Discrete Time Systems

Yoshihiro Yamamoto
Department of Information and Knowledge Engineering, Tottori University, Tottori, Japan
(Tel : +81-857-31-5624; E-mail: [email protected])

Abstract: This paper treats a new method for control design, called a simple model matching (SMM), not only for
linear, but also for nonlinear discrete time systems with/without time lags. This method is different with the
conventional model matching (MM) which is restricted to linear systems. SMM is the one which makes the system
output coincides with the desired output in one step. Conventional MM is, on the other hand, to make a closed loop
system matching with the desired system. Some features of the SMM are (1) It is very simple to be derived, (2) It is
useful for linear and nonlinear discrete time systems with/without time lags, (3) Decoupling control is easily realized for
multivariable systems. From these advantages described above, the SMM can be recognized as a unified approach for
discrete time systems. Multivariable case of this method is presented in this paper with some additional consideration
on time lag. Integral action is also discussed in detail with some simulation studies.

Keywords: Simple model matching, Nonlinear discrete time system, Time lag, Decoupling control.

1. INTRODUCTION y (k + 1) = ru (k ) + f k ( y , u ) (1)
It is well known that the Conventional Model and the desired system
Matching (MM) method [1] is one of the techniques for y d (k + 1) = f d ( y d (k ), y d (k −1), ..., u d (k ), u d ( k −1),...) (2)
control system design. Closed loop system is designed where y and u are an output and an input of the system,
to match its characteristic with the one of the desired and yd and ud are a desired output and a desired input
model by MM with some additional consideration for (reference input), respectively. Here, the abbreviation
robustness [2]. But, this method is, unfortunately, f k ( y, u ) ≡ f ( y (k ), y (k − 1),..., u (k − 1), u (k − 2),...)
restricted to linear systems, although some trial [3] has is used. The assumption posed on the system is only the
been done to extend the method for nonlinear systems. existence of a linear part of the current input variable.
A method, called Simple Model Matching (SMM), The purpose of the design is to determine the control
has been presented by the authors [4]. This is a method input so as to follow the system output to the desired
to determine a control input to attain a desired output in system output. Conventional model matching method is
one step which is simply derived by replacing the output the one to match a closed loop system with a desired
of the system with the desired output. Then, the system model. On the other hand, the simple model matching
output coincides with the desired output in one step. method is the one for the system output to match with
This fact holds for any nonlinear systems. Only the the desired model output in one step. This is simply
assumption posed on the system is that there exists a established by choosing a control input as
linear term of a current input variable. This is extended u (k ) = r −1{ y d (k + 1) − f k ( y, u )} (3)
to a system with time lag. A multivariable system is also
treated in a same manner, but it is important to note here which is derived by replacing the left hand side of (1)
that a multivariable system may include a time lag even with the desired output yd(k+1). By using this control
if the description of the system as it stands does not (3) to the system (1), it is easily shown that the system
explicitly contain a time lag. Furthermore, a decoupling output y(k+1) coincides with the desired output yd(k+1)
control system is easily realized if the desired system for any value of y(k). Fig.1 shows the typical responses
has a decoupled form. Some considerations for integral yd
1
action are also presented to cope with real situations SMM
where the system model does not coincide with the real 0.8 MM
system. 0.6
The essence of SMM method is described in Section2
0.4
for a scalar system. This is extended for a multivariable
Output

system in Section3 and integral action is presented in 0.2

Section4. Section 5 shows some simulation studies. 0

-0.2
2. SCALAR SYSTEM
-0.4

2.1 Nonlinear system without time lag -0.6

0 5 10 15 20
Consider the following nonlinear system Step
Fig.1 Outputs of SMM and MM

2325
of SMM and MM control. For the desired system fixed, y4 (k ) = y3 (k ) − cy1 (k ) − dy2 (k ) , (14)
the system output using the SMM control coincides with the part of the system which implicitly includes a time
the desired output in one step. On the other hand, the lag is appeared as follows
system output response approaches the desired output
y4 (k + 1) = q1T u(k − 1) + qT2 u(k − 2) + f 4 (k ) , (15)
asymptotically in the case of conventional MM control.
where q1 = s 3 − cs1 − ds 2 , q 2 = t 3 − ct1 − dt 2 , and f 4 = f 3
2.2 Nonlinear system with time lag −cf1 − df 2 with some required substitutions. Then, the
Consider the system with one step time lag overall system is represented by
y (k + 1) = ru (k − 1) + f k ( y, u ) (4) ⎛ y1 (k + 1) ⎞ ⎛⎜ r1 ⎞⎟
T ⎛ s1T ⎞
⎜ ⎟ ⎜ ⎟
and the desired system ⎜ T ⎟ ⎜ sT ⎟u( k − 1)
y
⎜ 2 ( k + 1) ⎟ = r2 u ( k ) +
yd (k + 1) = f d ( yd (k ), yd ( k − 1), ..., ud (k − 1),...) (5) ⎜ y (k + 2 ) ⎟ ⎜ ⎟ ⎜ 2 ⎟
⎝ 4 ⎜
⎠ ⎝ q1 ⎠
T ⎟ ⎜ qT2 ⎟
System (4) is also represented by ⎝ ⎠ (16)
y (k + 2) = ru (k ) + f k +1 ( y, u ) (6) ⎛ t1T ⎞
⎜ ⎟ ⎛ 1 ⎞
f ( k )
Then, the control input is seemed to be determined by ⎜ ⎟
+ ⎜ t t2 ⎟u(k − 2) + ⎜ f 2 (k ) ⎟.
u (k ) = r −1{ yd (k + 2) − f k +1 ( y, u )} (7) ⎜ ⎟ ⎜ f (k ) ⎟
⎜0 ⎟ ⎝ 4 ⎠
But, since fk+1(y, u) includes a future value of the system ⎝ ⎠
output, y(k+1), control input (7) is rewritten, in a [Cas2-1] rankR1=rank(r1, r2, q1)=3:
realizable form, by using SMM control is then obtained by
f k +1 ( y, u) ≡ f ( y(k + 1), y(k ),..., u(k − 1),...)
u(k ) = R1−1{y d (*) − S1u(k − 1) − T1u(k − 2) − f1 (k )} (17)
= f (r1u(k − 1) + f k ( y, u), y(k ), ..., u(k − 1),...) . (8)
where yd(*)=(yd1(k+1), yd2(k+1), yd4(k+2))T, S1=(s1, s2,
This argument can be extended for the case of time lag q2)T, T1=(t1, t2, q3)T and f1(k)=(f1(k), f2(k), f4(k))T are
of multi steps. defined with corresponding changes for the desired
model.
3. MULTIVARIABLE SYSTEM [Cas2-2] rankR1=rank(r1, r2, q1)=2:
In this case, q1 is also represented by q1=c1r1+d1r2 for
3.1 Nonlinear system without time lag
some c1 and d1. Then, similarly as above, a new variable
Consider the following nonlinear system y5 (k + 1) = qT3 u(k − 1) + qT4 u(k − 2) + f 5 (k ) (18)
y (k + 1) = Ru(k ) + f (k ) . (9) where q 3 = q 2 − c1s1 − d1s 2 and so on, leads to a new
This is a vector-matrix notation of n equations aspect of the system with a time lag of two steps. The
n
SMM control can be determined as in [Case2-1] if the
yi (k + 1) = ∑ r u (k ) + f (k )
j =1
ij j i i=1,2,…,n (10)
coefficient matrix of u(k) is non-singular. If this matrix
is singular, similar arguments continue until the leading
fi (k ) = fi ( y1 (k ), .., yn (k ), ..., u1 (k − 1),...,un (k − 1),..) . coefficient matrix of input variable has full rank.
If the square matrix R is nonsingular, then the SMM [Case 3] rankR=rank(r1, r2, r3)=1:
control is determined as In this case, there exist c and d such that r2=cr1, r3=dr1.
u(k ) = R −1{y d (k + 1) − f (k )} . (11) Then, as in the [Case2], this system has two subsystems
which include a time lag and the overall system is
This is the case where the system does not include any
represented by
time lag and can be treated in a way similar to the scalar
case. ⎛ y1 (k + 1) ⎞ ⎛⎜ r1 ⎞⎟
T ⎛ s 1T ⎞
⎜ ⎟ ⎜ ⎟
⎜ y 4 (k + 2) ⎟ = ⎜ p 1 ⎟u(k ) + ⎜⎜ p 2 ⎟⎟u(k − 1)
T T
3.2 Nonlinear system with time lag ⎜ y ( k + 2) ⎟ ⎜ ⎟
It is astonishing that the system (9) may contain a time ⎝ 5 ⎠ ⎜⎝ q 1 ⎟⎠
T ⎜ q T2 ⎟
⎝ ⎠ (19)
lag, even if the matrix R does not contain nonzero
⎛ t 1T ⎞ ⎛ f 1 (k ) ⎞
element. In order to make the discussion simple, ⎜ ⎟ ⎜ ⎟
consider the 3 variable case as follows. ⎜ ⎟
+ 0 u(k − 2) + ⎜ f 4 (k ) ⎟.
⎜ ⎟ ⎜ f (k ) ⎟
yi (k + 1) = riT u(k ) + sTi u(k − 1) + tTi u(k − 2) + f i (k ) . (12) ⎜0 ⎟ ⎝ 5 ⎠
⎝ ⎠
These are also presented by with appropriate changes of variables.
y (k + 1) = Ru(k ) + Su(k − 1) + Tu(k − 2) + f (k ) . (13) [Case3-1] rankR=rank(r1, p1 q1)=3:
[Case 1] rankR=rank(r1, r2, r3)=3: This is the case where two subsystems include one step
This is the case of 3.1 and the SMM control can be time lag and SMM control is given by the form similar
derived. to (10). If this is not the case, analogous discussions to
[Case 2] rankR=rank(r1, r2, r3)=2: [Case2-1] and [Case2-2] are necessary.
In this case, it is possible to assume that rank[r1, r2]=2 [Case 4] rankR=rank(r1, r2, r3)=0:
without loss of generality. Then, there exist scalar c and This case is trivial to include a time lag since r1=
d such that r3=cr1+dr2. Then, introducing a new r2=r3=0, i.e., R=0. Then the discussion is focused on S
variable and start with [Case1] using S instead of R.

2326
 u (k ) = rˆ −1{ yd (k + 1) − fˆk ( y, u )
3.3 Decoupling control (26)
+ q1 ( yˆ (k ) − y (k )) + q2 (Δyˆ (k ) − Δy (k ))}
It is almost clear that the decoupling control is easily
realized using the SMM control if the desired system Δyˆ (k ) = yˆ (k ) − yˆ (k − 1), Δy (k ) = y (k ) − y (k − 1)
has a decoupled form, i.e., (Proof) Control signal ( ) is equivalent to
yi , d (k + 1) = fi , d ( yd (k ), ..., ui , d (k ), ...) , i=1,2,…,n. (20) rˆu (k ) = yd (k + 1) − fˆk ( y, u )
+ q1 (rˆu (k − 1) + fˆk −1 ( y , u )) − q1 y (k )
4. INTEGRAL ACTION (27)
+ q (rˆu (k − 1) + fˆ ( y , u ))
2 k −1
In the preceding sections, it is assumed that the
system model is the same exactly with the system itself. − q2 (rˆu (k − 2) + fˆk − 2 ( y , u )) − q2 ( y (k ) − y (k − 1))
To cope with real situations where there exists a model This is arranged to
error, integral action is discussed in this section with rˆ(u (k ) − q1u (k − 1)) = q2rˆ(u (k − 1) − u (k − 2))
some simulation results to confirm the effectiveness of
the proposed method. − ( fˆ ( y , u ) − q fˆ ( y , u ))
k 1 k −1
(28)
+ q2 ( fˆk −1 ( y , u ) − fˆk − 2 ( y , u ))
4.1 Nonlinear system
− q2 ( y (k ) − y (k − 1)) + yd (k + 1) − q1 y (k )
For a system described by (1), system model is
assumed to be obtained in the form Here, set q1 = 1 and
yˆ (k + 1) = rˆu (k ) + fˆk ( y, u ) . (21) Δy = y (k + 1) − y (k ) = y (k ) − y (k − 1) ,
Then the control input is of the form Δu = u (k + 1) − u (k ) = u (k ) − u (k − 1) ,
u (k ) = rˆ −1{ y d (k + 1) − fˆk ( y, u )} . (22) Δf = fˆk ( y , u ) − fˆk −1 ( y, u ) = fˆk −1 ( y , u ) − fˆk − 2 ( y , u )
This causes the output error and especially, steady state for a large k in a rump response. Then,
error. (1 − q2 )rˆΔu = −(1 − q2 )Δfˆ − q2 Δy + yd (k + 1) − y (k ) (29)
4.2 Integral action follows and, if q2 = 1 ,
An integral action is easily introduced in the SMM yd (k + 1) − y (k ) = Δy = y (k + 1) − y (k ) . (30)
control as follows, Therefore, yd (k + 1) = y (k + 1) is concluded.
u (k ) = rˆ −1{ yd (k + 1) − fˆk ( y , u ) + q1 ( yˆ (k ) − y (k ))} . (23)
Here, the parameter q1 is introduced to join the case 5. SIMULATION STUDIES
with integral action ( q1 = 1 ) and the case without 5.1 Example-1
integral action ( q1 = 0 ). The case 0 < q1 < 1 is also This example shows the SMM control for nonlinear
useful to decrease the offset while keeping the closed system and the effectiveness of the power level of
loop stability. Therefore, the parameter q1 shows the integral action. System, system model, and desired
level of the integral action. model are represented as follows.
(Proof of the integral action) To prove that the term y (k + 1) = cu (k ) + cy (k ) /{1 + ( y (k − 1) − 1) 2 } ,
q1 ( yˆ (k ) − y (k )) has a role of the integral action, it is yˆ (k + 1) = u (k ) + y (k ) /{1 + ( y (k − 1) − 1) 2 } ,
assumed that the closed loop system is stable and the
yd (k + 1) = 0.3 yd (k ) + 0.18 yd (k − 1)
steady state of signals y (k ), yˆ (k ), y d (k ), u(k ) and
+ 0.3ud (k ) + 0.22ud (k ).
fˆk ( y , u ) are denoted by y , yˆ , yd , u and fˆ , c in the system is uncertain parameter of the system.
respectively. Then, substituting these value into (23), it Figs.2-5 are the ramp responses for c=1.6, 1 and 0.4.
follows that
rˆu = yd − fˆ + q1 ( yˆ − y ) , (24)
from which
y − y = (1 − q )( rˆu + fˆ − y ) = (1 − q )( yˆ − y )
d 1 1 (25)
Holds. Then, it follows that y (k ) = y d (k ) is concluded
in steady state if q1 = 1 .

4.3 Double integral action

Double integral action is also introduced without any
trouble as follows.

Fig.2 Ramp responses: q1 = q2 = 0

2327
 shows the case of q1 = 1 and q2 = 0.2 , which shows a
little better response than Fig.3. Therefore, the values of
q1 and q2 between 0 and 1 are sometime useful with
some tolerable offset.

5.2 Example-2
This is an example of two variable nonlinear system to
show the multivariable case and the decoupling control.
System:
y1 (k + 1) = sin y1 (k ) − 0.16 y1 (k − 1) − y2 (k ) + 0.8 y2 (k − 1)
+ 0.5u1 (k ) + 0.3u1 (k − 1) + 0.2u2 (k ) /(1 + 0.14u22 (k − 1))
Fig.3 Ramp responses: q1 = 1, q2 = 0
y2 (k + 1) = −0.92 y1(k ) + 0.544 y13 (k − 1) + y2 (k )
− 0.41 y2 (k − 1) + 0.1u1 (k ) + 0.02u1 (k − 1)
+ 0.4u2 (k ) + 0.08u2 (k − 1)
System model:
yˆ1 ( k + 1) = sin(1.2 y1 ( k )) − 0.176 y1( k − 1) − 1.2 y2 ( k )
+ 0.88 y2 ( k − 1) + 0.55u1 (k ) + 0.36u1 (k − 1)
+ 0.22u2 (k ) /(1 + 0.168u22 (k − 1))
yˆ 2 (k + 1) = −1.104 y1 ( k ) + 0.5984 y13 ( k − 1) + 1.2 y 2 ( k )
− 0.451 y2 ( k − 1) + 0.11u1 (k ) + 0.024u1 (k − 1)
+ 0.44u2 (k ) + 0.096u2 (k − 1)
Desired model:
Fig.4 Ramp responses: q1 = q2 = 1
yd 1 ( k + 1) = 0.9 yd 1 (k ) − 0.14 y d 1 (k − 1)
+ 0.3ud 1 (k ) + 0.18ud 1 (k − 1)
yd 2 (k + 1) = 0.8 yd 2 (k ) − 0.2 yd 2 (k − 1)
+ 0.5ud 2 (k ) − 0.1ud 2 (k − 1)
For the system, the system model and the desired model
described above, the SMM control is performed with
the reference input ud 1 (k ) ≡ 1 and ud 1 (k ) ≡ 1 for
testing the decoupling control and the initial condition
y1 (0) = 0.2 , y2 (0) = 0.4 and yd 1 (0) = yd 2 (0) = 0 .
Fig.6 shows the output responses of the SMM control
without integral action. There are some offsets since the
system model has some mismatching. If there is no
Fig.5 Ramp responses: q1 = 1, q2 = 0.2 mismatching and no difference on initial values, these
outputs coincide completely with the desired outputs. So,
The case c=1 implies the system model is exactly the the graph is omitted here in this case. On the other hand,
same as the system. Therefore, system output response Fig.7 is for the case with integral action. The errors have
coincides in one step with the desired model output
regardless of the values of q1 and q2 . Fig.2 is the 2.5

case without integration ( q1 = q2 = 0 ) and the output 2

errors become divergence for c=1.6 and 0.4. But the use 1.5

of integration ( q1 = 1, q2 = 0 ) suppresses these 1

divergence with some offset, which can be seen in Fig.3.
Output

0.5
To get the perfect matching, the double integral action
( q1 = q2 = 1 ) is used in Fig.4. The output of the case 0

c=1.6 is unfortunately unstable oscillation because of a -0.5

large model error, although the case of c=0.4 shows a -1
System 1
System 2
stable response as expected. In this case, it is possible to Desired Model 1
Desired Model 2
stabilize the system by adjusting the power level of the -1.5
0 10 20 30 40 50 60 70 80 90 100
integral actions, i.e., the value of q1 and q2 . Fig.5 Step
Fig.6 Decoupling control

2328
 2.5

1.5

1
Output

0.5

-0.5
System 1
-1 System 2
Desired Model 1
Desired Model 2
-1.5
0 10 20 30 40 50 60 70 80 90 100
Step
Fig.7 Decoupling control with integral action Fig.8 Output responses of example-3

been almost disappeared in spite of the difference of 6. CONCLUSIONS

initial values. Rigorously speaking, the decoupling
control is impossible if there is a model error. But the A simple model matching control method was
proposed method still attains the goal of decoupling presented for a multivariable nonlinear discrete time
control. system with/without a time lag. There is a case that the
system includes a time lag, even if it seems that there is
5.3 Example-3 no time lag on the exterior of the equations. The
coefficient matrix of the current input variable is
This is an example of two-variable case which includes
important to check the existence of a time lag.
a time lag implicitly.
Furthermore, a decoupling control can easily be
System:
executed using the proposed method.
y1(k + 1) = 0.3u1 (k ) − 0.21u1(k − 1) − 0.03u1 (k − 2) For a system model which includes a model error, an
+ 0.5u2 (k ) + 0.06u2 (k − 1) + 1.5 y1 (k ) − 0.66 y1 (k − 1) integral and a double integral actions were introduced to
+ 0.08 sin( y1 (k − 1)) − 0.5 y2 (k ) + 0.3 y2 (k − 1) cope with a steady state error. It was also stressed that
the power level of integral action introduced here was
y2 (k + 1) = 0.48u1 (k ) − 0.064u1 (k − 1) − 0.0144u1 (k − 2)
useful in some occasions.
+ 0.8u2 (k ) + 3.296u2 (k − 1) + 0.3 y1 (k ) − 0.6 y13 (k − 1) The proposed method is recognized as a unified
+ 1.9 y2 (k ) − 0.94 y2 (k − 1) + 0.12 y2 (k − 2) design method for nonlinear discrete time systems, since
this gives a design method for single and multivariable
Desired model:
systems with/without time lags in a same frame work.
yd 1 (k + 1) = 0.1ud 1 (k ) − 0.07ud 1 (k − 1) − 0.3ud 2 (k )
+ 0.9 yd 1 (k ) − 0.18 yd 1 (k − 1) REFERENCES
yd 2 ( k + 1) = 0.16ud 1( k ) − 0.48ud 2 (k ) − 0.08ud 2 (k − 1) [1] K. Ichikawa, Control System Design based on
+ 1.4 yd 2 (k ) − 0.48 yd 2 (k − 1) Exact Model Matching Techniques, Springer-
In this system, it seems that there is no time lag. But, in Verlag, Berlin, 1985.
the notation of (8), matrix R is of the form [2] Y. Yamamoto, “Model Matching and Adaptive
Control with Disturbance Consideration,” Trans.
⎛ 0.1 − 0.3 ⎞
R = ⎜⎜ ⎟⎟ . of SICE, Vol.24, No.11, 1210-1212, 1988.
⎝ 0.16 − 0.48 ⎠ [3] Y. Asada, et.al., “Synthesis of Nonlinear Model
Since rankR=1, the system includes a time lag. In order Matching Flight Control System for Tilt Rotor
to make a time lag appeared explicitly, new variables Aircraft,” Proc. of ICCAS2005, SA01-1, 2005.
y3 and yd 3 are introduced such that [4] Y. Yamamoto, et. al., “Simple Model Matching
y3 (k ) = y1 (k ) − 1.6 y2 (k ) , yd 3 (k ) = yd 1 (k ) − 1.6 yd 2 (k ) . Control Method for Nonlinear Discrete-Time
Systems,” Proc. of 6th APCCM, 190-194, 2004.
Representing the system and the desired model by these
new variables, a time lag is appeared explicitly in the
difference equation of y3 (k ) . The SMM control can be
applied for this system. As Fig.8 shows, the response of
the variable y1 (k ) which does not includes a time lag
coincides with the desired output in one step, but the
variable y3 (k ) including a time lag matches with the
desired output in one step after a time lag passed.

2329