Model-Based Multirate Controllers Design: Julián Salt and Pedro Albertos, Senior Member, IEEE
Model-Based Multirate Controllers Design: Julián Salt and Pedro Albertos, Senior Member, IEEE
Model-Based Multirate Controllers Design: Julián Salt and Pedro Albertos, Senior Member, IEEE
IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 13, NO. 6, NOVEMBER 2005
AbstractIn many industrial control applications the control action updating can be faster than the output measurement, leading
to multirate controllers. In this paper, some dual rate operations
are used to model the controller as well as the controlled plant.
The controller design is model-based and depends on the input to
be tracked. The controller is split into two parts acting at different
sampling rates and its design is approached based on the characteristics of each available sampling rate. The control target is to reach
the similar performances to those the faster single rate controller
would achieve. Model-based cancellation controllers are designed
using this approach and promising results are obtained.
Index TermsDigital control, modeling, model-based control,
multirate sampled data systems, proportionalintegral derivative
(PID) control.
NOMENCLATURE
CT
DT
DRDT
DRZOH
FSDT
MR
SSDT
Polynomial in
as defined in (5).
Continuous time.
Discrete time.
Dual rate discrete time.
Dual rate zero order hold.
Fast sampling discrete time.
DT transfer function of a CT plant with sample and
hold devices at -period.
Dual rate operator (input
-expanded sequence
and output -sequence).
DT controller designed to match the -period DT
model
.
Multirate.
Closed-loop transfer function of the plant with
controller.
Slow sampling discrete time.
Staircase signal from
.
Expanded polynomial (9).
Output of a DR controlled system.
sequence.
Z-transform of
sequence.
output of a plant with
controller.
-sequence expanded from an
-one.
-sequence obtained from a -one.
-unit operator
.
Manuscript received March 25, 2004; revised December 3, 2004. Manuscript received in final form July 29, 2005. Recommended by Associate
Editor K. Schlacher. This work was supported by the Plan Nacional de I+D,
Comisin Interministerial de Ciencia y Tecnologa (CICyT) under Program
DPI2003-01964.
The authors are with the Department of Systems Engineering and Control Universidad Politcnica de Valencia, Valencia 46022, Spain (e-mail:
julian@isa.upv.es; pedro@isa.upv.es).
Digital Object Identifier 10.1109/TCST.2005.857410
I. INTRODUCTION
Authorized licensed use limited to: IEEE Xplore. Downloaded on January 14, 2009 at 07:20 from IEEE Xplore. Restrictions apply.
But the main difficulty is the design of MR control algorithms. Many approaches have been reported in the literature
[15], [16]. One simple and direct approach is proposed in [17]:
Given a continuous time (CT) controller and a fast discrete time
(DT) controller obtained by discretization, a DT controller is redesigned by using a dual-rate hold device.
If a model of the process is available, some other alternatives are possible. Most of them are based on the feedback of
a computed output, which is estimated based on the process
model. There are many options to get the computed output, as
reviewed in [18]. Any time a measurement is available, the feedback signal could be the actual output, like in the inferential control, [19]. But, in a more general case, the computed output is
a function of both, the actual output, if it is available, and the
estimated output.
The approach in this paper is to design a two steps controller
with a slow part computing the control error anytime an actual
measurement is available, and generating, also using a model
of the process, a fast control updating to get results similar to
those achievable with a faster controller. This reduces the computing load to evaluate the complete control algorithm. Now a
slow part (remote control in a networked control system environment with shared medium) is assumed complementing a fast
one (local control). In the case of the classical inferential control techniques, just a local fast control based on some estimated
output (only slow measurement is known) is considered.
The paper is structured as follows. In the next section a
number of input/output DT models are reviewed. Then, the
control problem is properly defined. Different controller design
techniques, such as PID or cancellation are considered. In any
case, the use of cancellation controllers is part of the design,
raising the issue of robustness of the controlled plant. A number
of examples illustrate the results.
989
(1b)
where the series on the right-hand side only present powers of
. Note that the physical meaning does not allow a
trivial substitution.
On these sequences, some operations can be defined, [20].
i) The expand (upsampling) operator creates a
-sequence, as follows:
from an
-sequence
(2a)
because a T-sequence
In this case it is correct to use
is actually obtained.
ii) The skip (downsampling) operator creates an
sequence from a -sequence, as follows:
(2b)
The skip operation applied to the -transform of a signal can
be obtained using the expression (3), due to [21]:
A. Signals
From a CT signal, such as
in Fig. 1, a sequence of data
,
,
is represented by
.A
taken at rate
hold device
, operating at time periods , delivers a staircase signal
, from the input
. Ideally, a sequence, denoted by
, can be attached to the input of the hold device.
, the previous sequences are related to signal
For
values at the same time instant. Their -transforms can be expressed by
(3)
Some known [20] skip-expand properties usually considered
in this work are
(4a)
(4b)
(4c)
(1a)
that is, a) the skip operation does not commute, and b) the expand operation does. The third property is a clear rule with an
Authorized licensed use limited to: IEEE Xplore. Downloaded on January 14, 2009 at 07:20 from IEEE Xplore. Restrictions apply.
990
IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 13, NO. 6, NOVEMBER 2005
easy proof that will be used in the following. If the signals considered in (1) were finite time signals, their Z-transforms will be
polynomials.
Assumption a) is required to avoid aliasing [15]. By assumption b), the notation is simplified.
The following useful relationship is easily derived:
B. Process
Assume the CT process in Fig. 1. For any pair of sequences
like those in (1), considered as process output and input, respectively, a DT transfer function of the process plus the hold device
can be written.
1) The fast sampling DT (FSDT) model is defined by
(5)
,
are polynomials in
. As previously
where
mentioned, following the notation in (1a), these polynomials
, realso represent finite sequences with -elements
, being
, and
. The
spectively, for
expand and skip operators, (2), can be also applied to polyno.
mials in
2) For the same process, a slow sampling DT (SSDT) model
can be similarly defined by
(6)
where
,
are polynomials in
and, math. A treatment similar to that expressed for
ematically, in
the skip and expand signal operations will be assumed for these
polynomials.
applied to a system is defined as the Z
The operator
transform of the -period discretized impulse response of that
system.
. That is
The FSDT transfer function poles are denoted by
(9)
that is
(10)
where
.
, are distributed into
The parameters of the numerator,
groups of coefficients in such a way that the sum of each of
these groups leads to the slow single-rate numerator coefficient.
So, the SSDT model can be derived from the FSDT model, [22].
In the following, the arguments are omitted if their interpretation
is clear.
From (10), it yields (every variable expressed in )
(11)
If a skip operation is applied to the -time sequences, that is,
resampling, the result is
doing an
(12a)
(7)
whereas the SSDT poles are
where
is
, but expressed by means of the -varisamable. The physical meaning is the consideration of an
pler at the process output, that is, a slow output
, such that
(12b)
(8)
Authorized licensed use limited to: IEEE Xplore. Downloaded on January 14, 2009 at 07:20 from IEEE Xplore. Restrictions apply.
991
(13)
where
. In this case, it is possible to obtain a
transfer function of the process plus the DRZOH device
(14)
because the expand operation commutes, (4b). Also, by using
(13)
(15)
Fig. 2.
(16)
where
describes the transfer function from an exto a fast output
.
panded slow input
Using a similar notation, the DRZOH operation, (13), is rep.
resented by
Clearly, if the output is sampled with period
, it is obtained (see [22] for details)
(19)
As the expand (upsampling) operation commutes, (4b), and
assuming some elementary block algebra considerations
(20)
(17)
Authorized licensed use limited to: IEEE Xplore. Downloaded on January 14, 2009 at 07:20 from IEEE Xplore. Restrictions apply.
992
IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 13, NO. 6, NOVEMBER 2005
part given by
part given by
and a fast
. That is
(22)
b) Moreover, if the skipped fast output
should be
the same that the slow single rate control-loop output
, then the parts of the controller should be
a fast part given by
or
(23a)
(21a)
a slow part given by
or
, the
None of them will match
, and thus, for
corresponding DT equivalents of
step reference changes, none of the output sequences will
match the CT response equivalent ones.
2) In this case, the single rate controllers will be
(23b)
a rate converter with the form
(23c)
or
(21b)
For a step reference change, the CT controlled system output
is denoted by
. Now, the controlled system discretized
output for step reference changes will coincide with the corresponding DT sequences taken from the initial CT system, i.e.,
;
. That is to say
(21c)
Our MR controller design goal is to achieve the fast controlled
system performances
based on the slow output measurement sequence. The MR controller is composed by three parts,
as depicted in Fig. 2. The low rate computed error is first proand passed through a rate concessed by the subcontroller
verter (upsampling block). The second and faster subcontroller,
, finally provides the control input to the plant. The result
can be expressed as follows.
, and
Main Result: Given a CT process
for the controlled system, assuming
a reference model
a control updating rate
, the output being sampled at rate
, Fig. 2, the following hold.
a) The output response of the dual rate controlled plant to the
, being
its -period discretized
reference input
representation, will match the fast single rate output response, if the dual-rate controller is composed of a slow
(24)
(25)
Authorized licensed use limited to: IEEE Xplore. Downloaded on January 14, 2009 at 07:20 from IEEE Xplore. Restrictions apply.
(29)
it can be formulated that
993
(35)
(30)
and it is possible to express
(31)
hence, part b) has been proved.
Comment 1) is evident from the definitions of
,
,
.
To check 2), first realize that the fast controller, either (22)
or (23a), cancels the process dynamics, but the numerators of
and
are not the same. Thus, the intersampling
ripple could appear depending of
numerators roots,
[24]. If the fast controller is taken as
(32)
does not cancel the numerthe dual-rate controller
ator of the process transfer function, avoiding the ripple, but the
does not hold.
matching
With respect to comment 3), if the reference is a unitary step,
then
(33)
For a ramp input, that is
given by
(34)
from
or simply
This could provide a way to obtain
to express a link condition between both rate transfer functions.
must be known or computed beforehand.
Evidently
Remark 4: In general, the stability margins of the DR controlled system are between those of the slow and fast single rate
discrete time schemes. Nevertheless, due to the cancellation features, the design strongly depends on the process model.
Remark 5: If the process is nonminimum phase, the cancellation of unstable pole-zero pairs must be avoided. Thus, the fast
part of the controller could be alternatively computed by (32). If
the slow part is maintained the output does not match the output
predicted by the closed loop transfer function. Nevertheless, if
is computed by using (35) the response will follow the
.
performance fixed by
Remark 6: The approach works for unstable processes, as
far as there is not pole cancellation in (22) and (23a). This is
avoided by using (32) instead.
Remark 7: The inputoutput stability of the designed controlled system is ensured because the output matches some given
references and the possible ripple can be avoided. On the other
hand, as any other control design approach based on (partial)
model cancellation, the robustness of the result should be analyzed. In general, [27], the system internal stability requires that
none of the four functions
the sensitivity (or output/output_disturbance transfer) function;
the control/reference transfer function;
the output/control_disturbance transfer
function;
the control/control_disturbance transfer
function;
should have unstable poles. In the MR setting, this is achieved
as far as the fast part of the controller is implemented by (32).
This result allows for a simple design of dual-rate controllers,
easy to be implemented. Also, the separation of the controller
Authorized licensed use limited to: IEEE Xplore. Downloaded on January 14, 2009 at 07:20 from IEEE Xplore. Restrictions apply.
994
IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 13, NO. 6, NOVEMBER 2005
Fig. 3.
PID control.
for:
,
proximation for
,
, is given by
Given a CT process
, a CT PID controller
is
designed. If a discretized PID is applied [24], the controlled
process dynamic performances are degraded if a slow sampling
period is chosen. Assume the DT behavior of the controlled
plant is acceptable for a sampling period , with the fast con, but too poor if it is taken as
,
. The question
troller
, is
in the MR setting is: if the input can be updated at rate
it possible to get similar performances if the output is measured
s?
any
Authorized licensed use limited to: IEEE Xplore. Downloaded on January 14, 2009 at 07:20 from IEEE Xplore. Restrictions apply.
995
(36)
To avoid the ripple, the fast part of the MR controller is computed as (32), that is
The excellent results are shown in Fig. 5, where the step responses for the PID controlled plant for a CT controller, the
and
s, as
single rate PID controllers for
well as the dual-rate model-based controlled plant are plotted
altogether. Using the perfect plant model, the initial part of the
dual rate controlled plant response follows that achieved by the
s. a new
fast rate controller. However, at time
measurement is taken and the response is improved.
Obviously, the full controller is much more complex than either of the two simply discretized single rate PID controllers. It
is also better than the slow rate cancellation controller design to
.
match the targeted discretized transfer function
Authorized licensed use limited to: IEEE Xplore. Downloaded on January 14, 2009 at 07:20 from IEEE Xplore. Restrictions apply.
996
IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 13, NO. 6, NOVEMBER 2005
For instance in a shared medium distributed control environment, it is possible to have a central rough slow control and
best fast tuning after the information framepartial control actionhas traveled across the transmission medium.
The stability of the dual rate controlled plant can be assured
as the design is based on matching the reference responses obtained by the fast and slow rate predesigned (and stabilizing)
controllers.
Another relevant issue is the robustness against model uncertainty. This has not being used as a requirement in the design approach and will depend on how the desired closed loop transfer
function is defined. Anyway, the cancellation part of the controller will suffer from this internal disturbance.
By this approach it will be possible to extend the dual-rate
control design to a controller which is process-model independent by adjusting its fast and slow parts like a lag and lead phase
controllers, respectively.
Robustness of the design and frequency tuning of the controller, are currently issues under research.
V. CONCLUSION
C. General Comments
In the proposed control design approach, the model of the
plant is assumed to be known, both in CT and the DT versions.
Thus, as previously stated, if the model of the plant is used for
implementing the controller, why not use the model to predict
the output? By means of an inferential control the fast model
can be used to estimate the intersample output, [19]. Of course,
this is an alternative approach leading, under ideal working conditions, to similar results. It depends very much on the specific
design to decide which result is better. However, if the computed
output is used, there is only one degree of freedom to design the
controller, just considering the faster sampling rate.
The new approach provides a better insight in the slow and
fast activities of the controller and it can be considered as a tool
to design a controller focusing on the requested frequency range.
Two different time scales have been considered with the consequent benefits in computational load and easy understanding.
Authorized licensed use limited to: IEEE Xplore. Downloaded on January 14, 2009 at 07:20 from IEEE Xplore. Restrictions apply.
997
Authorized licensed use limited to: IEEE Xplore. Downloaded on January 14, 2009 at 07:20 from IEEE Xplore. Restrictions apply.
Note: This service is not intended for secure transactions such as banking, social media, email, or purchasing. Use at your own risk. We assume no liability whatsoever for broken pages.
Alternative Proxies: