CONTROL OF SYSTEMS WITH

TIME-VARYING DELAYS

Raimo Ylinen, Nikolai Vatanski

Laboratory of Process Control and Automation

P.O.Box 6100, 02015 HUT, Finland

Abstract: The paper presents a model-based control design method linear systems
with time-varying delays. The design is based on pole placement of input-output
models. Furthermore, the design of an output feedback controller is divided into
two independent design tasks: a predictor design and a feedback design for non-
delayed output.

Keywords: time-varying system, delay, skew polynomials, observer, controller

1. INTRODUCTION

A typical networked control system is described

in Fig.1. The main problem in their design is the
time-varying and often unknown delays between
the controlled system (process) and the controller
(N. Vatanski and Jämsä-Jounela, 2007). Even
though the delays can be measured the time-
variance makes the control design difficult because
of the need of time-varying design methods . In
this paper a design methodology based on time- Fig. 1. Control system with network delays
varying polynomial systems theory is presented
and applied to the observer and controller design.
equations over skew polynomials in an operator.
The consideration of delayed systems in continu- Then the pole placement designs of observers and
ous time with delay-differential models is compli- feedback controllers are considered (Blomberg and
cated and mathematically difficult. In particular, Ylinen, 1978; Ylinen, 1980; Blomberg and Yli-
this holds for time-varying systems. Instead, in nen, 1983).
discrete time the models can be presented, at
The methodology is applied to the networked
least as a good approximation, with time-varying
control system in Fig. 1. First, the overall model
difference equations. This results in a simpler
of the controlled system with input and output
methodology, even though the lack of continuity
delays is constructed. Then the observers for pre-
of signals and parameters causes some difficul-
diction of the process output are considered and
ties. In spite of these, the discrete-time models
a Smith predictor type predictor is taken as an
are used in what follows. First some basic con-
example. Finally, the feedbacks of the predicted
cepts and definitions of time-varying difference
output are parameterized.
systems and their interconnections as presented
in (Ylinen, 1975; Ylinen, 1980) are given. The Two numerical examples are presented. A sim-
mathematical descriptions are based on linear ple process is controlled over network with time-
varying delays. The operation of the time-varying
where rK = the unit delay operator on K and
Smith predictor is simulated both in open loop
and in closed loop with a time-invariant PI- 0K = the zero operator on K.
controller. Similarly, the use of the (unit) prediction operator
gives the skew polynomial ring K[q; qK , 0K ].

2. TIME-VARYING LINEAR DIFFERENCE Most of the concepts and properties of ordinary

SYSTEMS polynomials can be applied to skew polynomials.
Let X be ‘sufficiently rich’ to make the powers
2.1 Description of systems r0 , r1 , r2 , . . . linearly independent over K. Then
the representation of a skew polynomial a(r) is
Time-varying linear discrete time input-output unique and its degree deg a(r) is well-defined.
systems are usually described by difference equa- The choice X = CZ obviously guarantees this.
tions of the form The same choice for coefficients K, unfortunately,
leads to such kind of weak algebraic structures

n
m
ai (k)y(k − i) = bi (k)u(k − i) (1) which do not offer any methodology and tools for
i=0 i=0
consideration of system models above.
For instance, the division algorithms, e.g. the

where k ∈ Z = time set , u,y ∈ X = signal space (right) division algorithm (RDA)

⊂ CZ and ai ,bi ∈ K = coefficient space ⊂ CZ . a(r) = b(r)c(r) + d(r), deg d(r) < deg b(r) (9)
C, Z above denote the complex numbers and the
integers, respectively, and CZ is the set of infinite are satisfied uniquely for all a(r), b(r) = 0 if
bisequences over C. and only if the coefficient ring K is a field. This
Provided that the signal space is closed with is important because the division algorithms are
respect to the (unit) delay operator r needed for manipulation of skew polynomial ma-
trices used in descriptions of multivariable sys-
(rx)(k) = x(k − 1) (2) tems. Usual coefficient rings are not fields, but of-
ten they can be extended to their fields of fractions
and to the pointwise multiplication by coefficients and the signals to corresponding rational signals.
the equation ( 15) can be written as an operator
Note that a skew polynomial can be invertible as
equation
a skew polynomial only if it is of degree zero but

ai ri y = bi r i u (3) there can exist skew polynomials of higher degree
which are invertible as mappings (e.g r and q).
a(r) b(r)
A matrix with skew polynomial entries, i.e. a
skew polynomial matrix is unimodular if it is
Alternatively, the (unit) prediction operator
invertible as a skew polynomial matrix. Note
(qx)(k) = x(k + 1) (4) that for skew polynomial matrices there is no
determinant which could be used for testing the
can be used leading to the model unimodularity. Furthermore, a skew polynomial
matrix can be invertible as a mapping even though
ã(q)y = b̃(q)u (5) it is not unimodular.
Two skew polynomial matrices A(r), B(r) are row
Note that in the case X = CZ the operators r and
(column) equivalent if there is a unimodular ma-
q are invertible and q −1 = r.
trix P (r) such that A(r) = P (r)B(r) (A(r) =
Under B(r)P (r)). Skew polynomial matrices can be
some additional assumptions the opera-
tors ci ri constitute the (non commutative) ring brought to row or column equivalent forms e.g.
K[r; rK , 0K ] of skew polynomials (or skew polyno- to an upper triangular form using the elementary
mial forms) with respect to addition operations. These are: (i)add a row (column) mul-
tiplied from the left (right) by a skew polynomial
ai ri + bi r i = (ai + bi )ri (6) to another row (column), (ii)interchange of two
rows (columns), (iii)multiply a row (column) from
and multiplication the left (right) by an invertible skew polynomial.

( ai ri )( bi r i ) = ci ri (7)
2.2 Systems and compositions
which can be constructed by
A set of linear, time-varying difference equations
rb = rK (b) + 0(b) (8) can be written as matrix equations
3. ANALYSIS OF SYSTEMS AND
COMPOSITIONS

3.1 Systems

Realizability of a system model requires that the

output of the model can be solved uniquely when-
ever the input and a sufficient, finite number of
initial values of output are given. Then the model
is said to regular.
If only past and and present values of the input are
Fig. 2. General composition needed for solving the output, then the model is
nonanticipative or causal. If the model is presented
using the unit prediction operator q, then the
. y model is causal if and only if it is proper. For a
A(r)y = B(r)u ⇔ [A(r).. − B(r)] = 0 (10)
u single-input-single-output (SISO) system (5) this
means that the degree of ã(q) is not lower than
the degree of b̃(q).
where u ∈ X r ,y ∈ X s and A(r), B(r) are
skew polynomial matrices. Then the multivariable .
An IO-relation S generated by [A(r).. − B(r)] is
input-output (IO-) relation generated by (10) is said to be stable if every solution y to A(r)y = 0
defined as the set approaches 0 when the time t approaches the
infinity.
S = {(u, y)|A(r)y = B(r)u} (11)
It should be noted that in general the stability
cannot be tested from the ‘pointwise’ roots of
. det A(k)(r), where A(k)(r) denotes the ordinary
The matrix [A(r).. − B(r)] is called a generator for
polynomial matrix obtained from A(r) by replac-
S. Generators for the same input-output relation
ing the coefficients by their values at time k.
are input-output (IO-) equivalent. Obviously, two
row equivalent generators are IO-equivalent, but Let S be generated by
also multiplication by a matrix invertible as a
. .
mapping gives an IO-equivalent generator. [A(r).. − B(r)] = L(r)[A1 (r).. − B1 (r)] (14)
A composition of input-output relations consists
of a set of input-output relations (‘subsystems’) or Now, if L(r) is not invertible, S contains modes
their generators, and a description of the intercon- related to L(r) which cannot be affected by the
nections between the (signals of the) subsystems. input u. This means that S is not controllable.
Every composition can be brought to the general
form of Fig. 2, where Si is the internal IO-relation
3.2 Compositions
and So the overall IO-relation generated by the
composition. Conversely, the composition is then
Consider the composition of Fig. 2 and suppose
said to be a decomposition of So . Decompositions
that the composition is regular, i.e. the internal
of the same IO-relation are input-output (IO-)
IO-relation is regular. The generator (12) can be
equivalent.
brought to upper triangular form
For the internal IO-relation Si it is always possible  
.
to construct a generator from the generators of the A 1 (r) A
2 (r) .. −B
1 (r)
  (15)
subsystems and the interconnections ..

0 A (r) . −B (r)
4
2
 
..
 A1 (r) A2 (r) . −B1 (r)  Now if for each (uo , yo ) satisfying the equation
.. (12)
A3 (r) A4 (r) . −B2 (r) Ã4 (r)yo = B̃2 (r)uo (16)

Instead, for the overall IO-relation there exists a y1 such that (uo , (y1 , yo )) satisfies
the equation
So = {(uo , yo )|∃y1 [(uo , (y1 , yo )) ∈ Si ]} (13) Ã1 (r)y1 = −Ã2 (r)yo + B̃1 (r)uo (17)

then the overall IO-relation So is generated by the

the construction of a generator is a more compli- .
cated task. equation (16) or by the generator [Ã (r)..− B̃ (r)].
4 2
If Ã1 (r) is invertible, then the y1 satisfying (17) polynomials are invertible mappings only if this
must be unique. In this case the composition is range is the whole X which means that each value
observable. If the system is causal, then it is always u(k − θ(k)) appears only once in the values v(k).
possible to take Ã1 (r) = I . Similarly to the input delay above, the output
Consider again the composition of Fig.2. If the delay can be written in two ways
generator of Si can be brought to the form
y = (f0 + f1 r + ... + fn rn )x = f (r)x (28)
 
.
1 (r) 0 .. −B
A 1 (r) (ẽ0 + ẽ1 q + ... + ẽn q n )y = ẽ(q)y = x (29)
  (18)
..

A (r) I . −B (r)
3 2

4.2 Overall system

the composition is called a generalized state space
decomposition of So , and y1 is the corresponding
Now, if the process is described by (3), the con-
generalized state.
trolled system is a series composition with internal
model generated by
4. SYSTEM WITH TIME-VARYING DELAYS  
..
 1 0 0 . −d(r) 
4.1 Time-varying delays  . 
 −b(r) a(r) 0 .. 0  (30)
 
.
Return to the system in Fig.1. Provided that 0 −f (r) 1 .. 0
the time-varying input delay is a multiple of the
sampling interval, it can be presented as with variables (v, x, y, u). Using elementary row
v(k) = u(k − θ(k)) = (r θ(.)
u)(k) (19) operations this can be brought to the form
 
..
This is not a skew polynomial representation but  1 0 0 . −d(r) 
it can be written as a skew polynomial equation  . 
 0 a(r) 0 .. −b(r)d(r)  (31)
 
v = (d0 + d1 r + ... + dn rn )u = d(r)u (20) ..
0 −f (r) 1 . 0
where the coefficients are zero otherwise but
so that v can be uniquely eliminated. Then the
dθ(k) (k) = 1 (21) remaining model is written using the operator q
 
..
Another way to describe the time-varying delay is  ã(q) 0 . −b̃d (q)  (32)
to use prediction .
−1 ẽ(q) .. 0
v(k + ϑ(k)) = (q ϑ(.) v)(k) = u(k) (22)
Using again the elementary row operations gives
which can be written as a skew polynomial equa-  
tion ..
 1 −ẽ(q) . 0 
.. (33)
(c̃0 + c̃1 q + ... + c̃n q n )v = c̃(q)v = u (23) 0 ã(q)ẽ(q) . −b̃ (q)
d

If the delay θ and the prediction interval ϑ are Thus x can be uniquely eliminated and the overall
related to each other by system is generated by
ϑ(k) = θ(k + ϑ(k)) (24)
.
ã(q)ẽ(q) .. −b̃d (q) (34)
then
q ϑ(.) rθ(.) u = u (25) 5. OBSERVER DESIGN

for all u ∈ X . On the other hand, if 5.1 General observer

θ(k) = ϑ(k − θ(k)) (26)
Consider the composition of Fig. 2 and suppose
then that only the overall input uo = u and output
yo = y2 are measured. The problem is to design a
rθ(.) q ϑ(.) v = v (27) dynamic system, a so-called observer for continu-
ous estimation of the internal output y1 , so that
but only if v belongs to the range of rθ(.) . Thus the estimation error ỹ1 = y1 − y1 behaves in a
rθ(.) and q ϑ(.) as well as the corresponding skew satisfactory way.
Let the internal IO-relation Si be generated by the
generator (15) of the upper triangular form and
the observer S to be designed by the generator
.
[C(r).. − D (r) − D (r)]. In what follows, (r) (or
1 2
(q)) is in some places omitted in order to shorten
the notations.
If the observer is chosen to satisfy

C −D1 −D2 T1 T2 1 A
A 2 −B
1
0 A 4 −B 2 = 0 I 0 A 2 (35)
4 −B

for some T1 , T2 then the error is generated by

Fig. 3. General feedback control composition
C ỹ = T1 Ã1 ỹ1 = 0 (36)
The input and output delays are θi and θ, respec-
tively, and corresponding predictions ϑi and ϑ.
(Ylinen, 1980; Blomberg and Ylinen, 1983). Thus
The resulting Smith predictor for x is
the design problem has been changed to the con-
struction of the matrices T1 , T2 . The matrix T1 x̂(k) = 0.99x̂(k − 1) + y(k − θ(k − 1)
affects the stability of the estimation error and +ϑ(k − θ(k − 1))) − 0.99y(k − 1)
after T1 of order high enough has been chosen the (42)
+0.01u(k − 1 − θi (k − 1)) − 0.01u(k − 1
matrix T2 is used to achieve a causal (proper, if q is −θ(k − 1) − θi (k − 1 − θ(k − 1)))
used) observer. Both matrices can be constructed
sequentially using the elementary row operations.

6. FEEDBACK COMPENSATOR DESIGN

5.2 Smith predictor
6.1 General feedback
The well-known Smith predictor is often used for
compensating delays in control loops (N. Vatanski
and Jämsä-Jounela, 2007). In the time-invariant Consider the feedback composition in Fig.3 con-
case the design is simple but for time-varying sisting of an IO-relation S to be compensated
systems more complicated. Consider the delayed and a feedback compensator S2 to be designed
system (30) and construct an observer for x using so that the resulting composition is stable, robust,
the input u and output y. realizable etc. Let S be controllable and generated
.
The design is started from (33). Multiplication by by [A..−B] and the feedback IO-relation S2 be gen-
.
erated by [C ..−D]. Then the feedback composition
ã(q) 1
(37) is generated by
0 1

A −B I 0 A −B
= (43)
results in the open observer −D C T3 T 4 Q3 Q4
 
.. Q
 ã(q) 0 . −b̃d (q) 
.. (38)
0 ã(q)ẽ(q) . −b̃ (q)
d where Q is invertible and can be constructed by el-
ementary column operations and T3 , T4 are appro-
Furthermore, multiplication of (33) by priate matrices (Blomberg and Ylinen, 1983; Yli-
nen, 1980). The dynamic behaviour of the system
ẽ(q)ã(q) ẽ(q) − 1 depends on T4 . Thus the feedback compensator
(39)
0 1 can be designed starting from a suitable T4 and
constructing then T3 so that the resulting feed-
gives the Smith predictor back compensator is causal and the whole compo-
  sition is robust against the parameter variations.
..
 ẽ(q)ã(q) −ã(q)ẽ(q) . −(ẽ(q) − 1)b̃d (q)  (40)
The construction can be carried out step by step
. using elementary row operations.
0 ã(q)ẽ(q) .. −b̃ (q)
d
The generalized state representations (18) can be
Example. Consider the system described by controlled by state feedback. Instead of the state
y1 the corresponding estimate y1 determined by
(1 − 0.99r)x = 0.01rv (41) the observer (35) can be used for feedback control.
Fig. 4. Input delay θi and output prediction ϑ = θ Fig. 5. Responses of PI controlled system
6.2 Control of system with time-varying delays where δ1 , δ2 , ..., δ6 are functions of delays. A simu-
lated responses of the closed loop system without
Consider again the system of Fig.1 and suppose and with the Smith predictor are presented in
that the non-delayed output x is available for feed- Fig.5.
back. Hence the design can be started from (31)
.
or from the generator [a(r).. − b(r)d(r)]. Suppose 7. CONCLUDING REMARKS
further that the system is controllable and the
time-varying delay such that d(r) is an invertible The time-varying polynomial systems theory gives
mapping. tools for the analysis and design of linear estima-
. tors and feedback controllers. In this paper the
Let the feedback controller be generated by [l(r)..− methodology has been applied to design of pre-
m(r)]. Then the feedback composition is gener- dictors and controllers for delayed systems with
ated by time-varying but measurable delays. The main
problems in the design are related to compli-
a −bd 1 0 a −bd
= (44) cated symbolic calculation of skew polynomials.
−m l t3 t4 q3 q4 d
For multivariable systems the calculation must be
done using special software for symbolic mathe-
Thus all possible feedback controllers can be gen- matics. Anyway, the theory gives tools to analyze
erated by the situation with respect to time-variance.
. .
[l.. − m] = [−t3 bd + t4 q4 d..t3 a + t4 q3 ] (45)
REFERENCES
varying the parameters t3 (r) and t4 (r).
Blomberg, H. and R. Ylinen (1978). Foundations
Example. Consider a second order system de- of the polynomial theory for linear systems.
scribed by (3) with Int. J. General Systems 4, 231–242.
Blomberg, H. and R. Ylinen (1983). Algebraic
theory for multivariable linear systems. Aca-
a(r) = 1 − 0.83r + 0.0055r2 (46)
demic Press.
2
b(r) = 0.15r + 0.033r N. Vatanski, J.-P. Georges, C. Aubrun E. Ron-
deau and S.-L. Jämsä-Jounela (2007). Net-
and with time-varying input delay θi and output
worked control with delay measurement and
prediction ϑ = θ presented in Fig. 4. The system
estimation. Control Engineering Practice,
is controlled by a PI controller
submitted.
(1 − r)u = (0.055 − 0.0098r)(yref − y) (47) Ylinen, R. (1975). On the algebraic theory of lin-
ear differential and difference systems with
time-varying or operator coefficients. Techni-
A Smith predictor is used to compensate the
cal report. Helsinki University of Technology,
output delay and it is generated by
Systems Theory Laboratory.
x̂(k) = 0.83x̂(k − 1) − 0.0055x̂(k − 2) Ylinen, R. (1980). An algebraic theory for analysis
+y(k − δ1 (k)) − 0.83y(k − 1 − δ2 (k)) and synthesis of time-varying linear differen-
+0.0055y(k − 2) + 0.15u(k − 1 − δ3 (k)) (48) tial systems. Acta Polytechnica Scandinavica,
+0.033u(k − 2 − δ4 (k)) − 0.15u(k − 1 − δ5 (k)) Mathematics and Computer Science Series.
−0.033u(k − 2 − δ6 (k))