You are on page 1of 20

AAS 05-327

THE EFFECT OF INTERPOLATION ON STABILITY AND


PERFORMANCE IN REPETITIVE CONTROL

Wondo Kang1 and Richard W. Longman2

Repetitive control (RC) can potentially eliminate the influence of vibrations on


fine pointing equipment caused by periodic disturbances, such as from slight
imbalance in a momentum wheel. Typical digital RC assumes that the
disturbance period is an integer number of time steps, nominally using the
nearest integer as the period which introduces substantial error at higher
frequencies. This paper studies the effect of introducing linear or cubic
interpolation to address this issue. In RC there is a heuristic monotonic decay
condition that is a sufficient condition for stability. It is also very close to being
a necessary condition. It is shown that RC designed based on this condition
without regard for interpolation, will also be stable with interpolation
introduced. Furthermore, interpolation may help stabilize the RC. Zero-phase
low-pass filtering is often introduced to robustify RC to residual modes, and it is
shown that a filter cutoff designed without considering interpolation, will also
stabilize the system with interpolation. The relationship to higher order RC is
investigated. Interpolation introduces steady state error for disturbances at the
addressed frequencies, and these errors are investigated as a function of
frequency. Finally, the influence of interpolation at other frequencies that are not
being addressed by the repetitive control design, are studied.

INTRODUCTION

Repetitive control (RC) is a form of control that specifically addresses the problem of
canceling the effects of periodic disturbances on the output of a feedback control system (Refs. 1-
18). For each time step, the approach looks at the error in the previous period or periods and
adjusts the command in order to converge on a command that produces zero tracking error as the
time steps progress. Spacecraft often have dynamic components such as a momentum wheel,
reaction wheels, control moment gyros, cryo pumps, etc. With fine pointing equipment onboard,
there can be a need to isolate the equipment from the periodic vibrations caused by a slight
imbalance in such components. Repetitive control has the potential to fully cancel the effects of
these periodic disturbances on fine pointing equipment.

Since the periods of these disturbances are determined by influences external to the
control system, there is no reason for the period to be an integer number of time steps.
Furthermore, in the case of reaction wheels the period gradually builds up over time. The usual

1
Graduate Student, Department of Mechanical Engineering, Columbia University, New York, NY 10027
2
Professor of Mechanical Engineering and of Civil Engineering, Columbia University, New York, NY
10027
repetitive control system, necessarily implemented in discrete time, assumes the disturbance
period is an integer number of steps, and in practice one would normally pick the nearest integer.
At any time step, repetitive control wants to look at the error one period back (and then forward
one time step to reflect the influence of the usual one time step delay going through the system) to
decide on the control update. In discrete time, one most likely does not have any data taken at the
precise time of interest. It is natural to introduce interpolation to address this issue. A separate
work, Ref. 10, studied some aspects of interpolation in repetitive control. Here, an overall
assessment of the influence of interpolation on the performance of repetitive control is made. One
might think that if one samples fast, this issue will not be important. But repetitive control aims
for zero error at every sample time, and aims to eliminate not just the fundamental of the chosen
period in the disturbance, but also all harmonics up to Nyquist frequency. The interpolation
decays rapidly in quality as the frequency increases. The purpose of this paper is to study the
effect of interpolation on stability, the effect of interpolation on the final error level reached, and
the influence of interpolation on the waterbed effect. The waterbed effect is a fundamental
property of linear feedback systems, that if errors from disturbances in some frequencies are
substantially attenuated by the feedback system, then errors at other frequencies must be
amplified. In addition, this paper studies the relationship of interpolation and higher order
repetitive control.

MATHEMATICAL FORMULATION OF FIRST ORDER RC

Figure 1 shows the block diagram of a typical repetitive control system where YD (z) is
the desired periodic or constant output, E(z) is the measured error, R(z) is the repetitive
controller transfer function, U(z) is the command to a feedback control system that is adjusted
every time step by the repetitive controller, G(z) is the closed loop transfer function of the
!
feedback control system, V (z) is a deterministic periodic disturbance, and Y (z) is the output. A
! !
periodic disturbance might enter somewhere within the feedback control loop, but there is always
! output disturbance which is used here. Repetitive control aims to eliminate
an equivalent periodic
! with a constant command, or eliminate the tracking error
the influence of a periodic disturbance
!
following a periodic command without any periodic disturbance, or ! eliminate errors in tracking a
periodic command and errors from a periodic disturbance, simultaneously, when they both have
the same period.

In this section, we temporarily consider that the period is an integer number of time steps,
and develop the governing equations. Then, in following sections we parallel this development
for different situations involving interpolation. Let pT be the period of the periodic disturbance
or command, where T is the time interval between sample times, and p is an integer. The simplest
form of repetitive control can be written as
! " p) + #e(k " p +1)
u(k) = u(k (1)

where k is the current time step, " is a repetitive control gain, and the sample time interval T is
dropped from the notation for simplicity. This repetitive control law can be interpreted as follows:
!
if the output was two units too small at the current phase but in the previous period, then add two
units times the chosen gain " to the command to the feedback controller this period. A one time
!
step lead is introduced to account for the usual one time step delay going through a feedback
control system (and the lead must be adjusted accordingly if the delay is a different number). It is
!
convenient to design the repetitive control law in the frequency domain using a more general
form

z pU(z) = F(z)[U(z) + z"(z)E(z)] (2)

This generalizes Eq. (1) by replacing " by "(z) , which can include not only a gain, but also a
compensator. Also, F(z) is introduced which can be a zero-phase low-pass filter used to make a
!
frequency cutoff of the learning process. Such a cutoff makes it easier to obtain a convergent
process, but at the expense of not trying to learn to eliminate periodic errors above the cutoff
!
! effective
frequency. Reference 8 presents methods to design "(z) and Ref. 9 presents effective
!
methods to design F(z) (see also Refs. 5-7). Then the repetitive control law transfer function
becomes
!
U(z) F(z)z"(z)
! R(z) = = p (3)
E(z) z # F(z)

and the relationship of the command and the periodic disturbance to the error can be written as
! [1+ G(z) R(z)]E(z) = YD (z) "V (z) (4)
{z " F(z)[1" z#(z)G(z)]}E(z) = [z p " F(z)] [YD (z) "V (z)]
p
(5)

There are several ways ! to interpret the last equation, which can be thought of a difference
equation. Consider first, the right hand side forcing function. If there is no filter F(z) so that it is
replaced !by the number 1, then the right hand side becomes zero because it represents the
difference of periodic functions, examined one period ahead of the present time and the present
time. If one had to approximate the period by the nearest integer value of p, then the right hand
side would not be zero, and there would be a forcing function. Figure ! 4, which is included in the
paper in a different context discussed below, shows how the magnitude of the frequency transfer
function of z p "1 on the right, varies with frequency from zero to Nyquist frequency. We see that
when the period is not perfect, one will not get zero. The nearest integer would not normally
produce the worst case, but the worst case has the right hand side doubling the amplitude of the
disturbance. This case can apply when there are additional disturbances at frequencies other than
! addressed by the repetitive controller. Now consider the left hand side. When the right hand
those
side is zero so that there is no forcing function, the tracking error will converge to zero as time
step k increases, for all initial conditions, if and only if all roots of the characteristic polynomial

z p " F(z)[1" z#(z)G(z)] = 0 (6)

are inside the unit circle. Starting from Eq. (5) there is a heuristic interpretation of stability.
Stability is purely a function of the homogeneous equation, and it does not matter whether there is
a forcing function on the! right or not. So we consider the homogeneous equation and rewrite it as

z p E(z) = F(z)[1" z#(z)G(z)]E(z) (7)

By setting z = e i"T in the square bracket in Eq. (7), the transfer function in brackets appears to be
the frequency response transfer function from one repetition to the next. This makes a quasi
steady-state assumption! for each period. The equation suggests that requiring

!
F(e i"T )[1# e i"T $(e i"T )G(e i"T )] <1 (8)

for all ! up to Nyquist frequency, would establish convergence to zero tracking error, with the
amplitude of each frequency component being multiplied by a factor less than one going from
! Reference 7 proves that satisfying Eq. (8) is a sufficient condition for
one period to the next.
stability, and the arguments used in that reference will be paralleled in a later section in order to
include interpolation in the proof. It is also very close to being a necessary condition for values of
p greater than roughly 10 (Ref. 11). As a result, it is very useful to use Eq. (8) to design repetitive
controllers, both to ensure stability, and as a way of aiming for well behaved monotonic decay of
the tracking error. We will examine the question, if one designs an RC system using Eq. (8), and
then needs to use interpolation, can the interpolation destabilize the repetitive control process?

INTRODUCING INTERPOLATION INTO REPETITIVE CONTROL

Linear Interpolation

When the period pT does not correspond to an integer value for p, one can introduce
linear interpolation into Eq. (1) to approximate an unknown such as e(k " p +1) in terms of the
known quantities e(k " p" +1) and e(k " p + +1)

u(k) = a1[u(k " p" ) + #e(k " p" +1)] + a 2 [u(k "!p + ) + #e(k " p + +1)] (9)
! !
where p + is the nearest integer of the period p (given in units of times steps) on the positive side,
"
and p! is the nearest integer of the period p on the negative side, and a1 = p + " p and a1 + a 2 =1 .
More generally, we write this law in z-transform space and introduce the compensator as above
! +!
z p U(z) = I 2 (z)[U(z) + z"(z)E(z)] (10)
! ! ! !
I 2 (z) = a1z + a 2

Temporarily, we ignore the possibility of using a zero phase filter F(z) , and re-introduce it later.
!
The repetitive control law transfer function with linear interpolation is then
!
U(z) I 2 (z)z"(z)
R(z) = = + ! (11)
E(z) z p # I 2 (z)

The analogous equation to Eq. (5) is


+ ! +
{z p " I 2 (z)[1" z#(z)G(z)]}E(z) = [z p " I 2 (z)] [YD (z) "V (z)] (12)

and we see that the interpolation transfer function appears in the same location as the zero-phase
low-pass filter in Eq. (5). When there is no interpolation, I 2 (z) is replaced by the number 1. Once
!
linear interpolation is introduced, the difference over one period on the right will no longer be
perfectly zero, there will be a forcing function due to error due to the imperfect nature of
interpolation. One expects this error to be smaller than the error discussed above related to using
the nearest integer for p instead of interpolating.!Convergence of the repetitive control with linear
interpolation as time step k goes to infinity, for all initial conditions, will ask that all roots of the
characteristic polynomial
+

! z p " I 2 (z)[1" z#(z)G(z)] = 0 (13)

are inside the unit circle of the z -plane. This time the process will converge to the steady state
particular solution associated with the non-zero forcing function on the right hand side of Eq.
(10). The heuristic quasi! steady-state condition for stability and monotonic decay of the error,
analogous to Eqs. (7) and (8), becomes
+
z p E(z) = I 2 (z)[1" z#(z)G(z)]E(z) (14)
i"T i"T i"T i"T
I 2 (e )[1# e $(e )G(e )] <1 (15)

One asks that Eq. (15)! be satisfied for all " up to Nyquist, and we will show below that this
guarantees convergence to the steady state particular solution associated with the forcing function
!
on the right in Eq. (12).
!
Cubic Interpolation

As the frequency goes up, a linear interpolation of a sinusoid gives an increasingly


inaccurate estimate of the value of the sinusoid at intermediate points. By using a higher order
interpolation scheme one could get improved interpolation results, at least for a low to middle
range of frequencies. Consider using two points behind and two points ahead of the point of
interest, making a cubic interpolation. Then

u(k) = u (k) + "e (k +1) (16)


u (k) = a1u(k " p "1) + a2 u(k " p + ) + a 3 u(k " p + +1) + a 4 u(k " p + + 2)
+
(17)
" = p + # p, $ =1# "
! a1 = #(1/6)"$ (1+ $ )
! a2 = (1/2)(1+ " ) $ (1+ $ )
a3 = (1/2)(1+ " )" (1+ $ )
a4 = #(1/6)(1+ " )"$

with e (k +1) defined analogously. In transfer function form with the compensator introduced
! +
z p U(z) = I 4 (z)[U(z) + z"(z)E(z)] (18)
! I 4 (z) = a1z "1 + a 2 + a3 z + a 4 z 2

The z-transfer function of the repetitive controller, the difference equation, the characteristic
!
polynomial, and the heuristic monotonic decay condition are as in Eqs. (11) through (15), with
I 2 (z) replaced by I 4 (z) !
.

! !
Using Zero-Phase Low-Pass Filtering Together with Interpolation

The usual stability condition is difficult to satisfy. And even if one can satisfy it all the
way to Nyquist frequency, based on a nominal model, it is very easily destabilized in practice due
to model error at high frequency. All that is needed is that the model be missing some residual
vibration mode or pole that contributes roughly 90 degrees or more phase lag. In order to produce
robustness to unmodeled high frequency modes, one can use a real-time zero-phase low-pass
filter on the entire control signal. This means that one no longer tries to fix errors from harmonics
in the disturbance above the cutoff. An important question is, if one has established the needed
cutoff frequency without any consideration of interpolation, can introducing interpolation force
one to adjust the cutoff? It is established that the answer is no, that interpolation will not
destabilize a repetitive control system that has been stabilized by this approach, making use of a
cutoff to succed in satisfying Eq. (8). In fact, if the interpolation needed does not change with
time, i.e. if the period is truly fixed, one might even be able to increase the cutoff frequency when
interpolating. The interpolation acts as an additional low pass filter (that is not zero phase). When
a zero-phase low-pass filter F(z) is introduced along with linear or cubic interpolation, the
update equation for the command to the feedback controller becomes

+ U(z) F(z)I j (z)z"(z)


z p U(z) = F(z)I
! j (z)[U(z) + z"(z)E(z)] : R(z) = = + (19)
E(z) z p # F(z)I j (z)

And the difference equation, the characteristic polynomial, and the heuristic monotonic decay
!condition become
!
+ +
{z p " F(z)I j (z)[1" z#(z)G(z)]}E(z) = [z p " F(z)I j (z)] [YD (z) "V (z)] (20)
p+
z " F(z)I j (z)[1" z#(z)G(z)] = 0 (21)
F(e i"T )I j (e i"T )[1# e i"T $(e i"T )G(e i"T )] <1 ; 0 % "T % & (22)
!
Equation (22) is the same
! as Eq. (8) except that attenuation at high frequencies is produced not
only by the zero-phase low pass filter, but also by the interpolation, which can be though of as a
! we will prove that Eq. (22) is a sufficient condition for stability. Let us examine in
filter. Below,
more detail how the interpolation term influences the inequality. Consider first order
interpolation. Then

(a1e i"T + a 2 )F(e i"T )[1# e i"T $(e i"T )G(e i"T )] = a1e i"T + a 2 F(e i"T )[1# e i"T $(e i"T )G(e i"T )]
(23)
Note that
! 1 1
a1e i"T + a 2 = (a12 + 2a1a2 cos "T + a22 ) 2 # [(a1 + a 2 ) 2 ] 2 =1 (24)

In other words, if a repetitive control system is designed to satisfy Eq. (8) for all frequencies up to
Nyquist, which ensures stability without interpolation as shown below, then introducing linear
!
interpolation will not destabilize the system. Figure 2 studies the magnitude of I 2 (e i"T ) for
0 " a1 "1. The plot uses the normalized frequency " = #T which starts from zero and reaches
Nyquist at " = # . The interpolation would be perfect if the plot equaled unity at all frequencies.
!
! !
!
When a1 = 0 it is perfect because the period is pT = p +T , and there is no need for interpolation.
The same is true for a1 =1 . The curves are the same for a1 and for 1" a1 , so all values of a1
between 0 and 1 are represented in increments of 0.1. The interpolation acts like a low pass filter,
attenuating the most for a frequency half way between integer values of period, and with no
! !
attenuation at integer values. Hence, interpolation mimics the behavior of the low pass filter, and
can produce!stability in systems that would otherwise
! not be!stable. For cubic interpolation,
! we
have no analytical proof that the magnitude of the interpolating filter satisfies

I 4 (e i"T ) #1 (25)

However, the corresponding magnitude plots are given in Figure 3. This time, " plays the same
part as a1 did for linear interpolation. Again, it appears that inequality (25) holds, and the same
!
conclusion as below Eq. (24) applies to cubic interpolation. Note that cubic interpolation stays
closer to unity up to higher frequencies, so that the forcing function!to the difference equation
will stay closer to zero, as desired for good performance.
!
The above considers that Eq. (8) is satisfied, and asks whether the RC system is still
satisfied if interpolation is introduced. Consider instead that we design a repetitive control system
to satisfy (22). The interpolation introduces an additional low pass filter, so that it is easier to
satisfy (23) than it is to satisfy (8). The difference between the two conditions depends on the
value of a1 for linear interpolation, and if it happens to be zero or unity, then there is no
difference. Hence, one can only take advantage of the less restrictive condition (23) if a1 has a
value that is not zero or unity, and in addition, one is confident that the period will never change.
We conclude that only in special situations would one want to create an RC system design to
!
satisfy (23) without satisfying (8).
!
INTERPOLATION IN HIGHER ORDER REPETITIVE CONTROL

Some have suggested the use of higher order repetitive control. This term is used for RC
that makes command updates to the feedback control system based on error in more than one
previous period (see for example, Refs. 12-16). Consider a second order RC law of the simplest
form, using error and input information from the previous two repetitions.

u(k) = "1u(k # p) + " 2 u(k # 2 p) + $["1e(k # p +1) + " 2e(k # 2 p +1)] (26)

In order to converge to zero error when the period p is an integer, the sum "1 + " 2 must be chosen
equal to unity. Now introduce interpolation to create
!
u(k) = "1[a1u(k # p# ) + b1u(k # p + )]
!
+ " 2 [a 2 u(k # (2 p) # ) + b2 u(k # (2 p) + )]
(271)
+ ${"1[a1e(k # p# +1) + b1e(k # p + +1)]
+ " 2 [a 2e(k # (2 p) # +1) + b2e(k # (2 p) + +1)]}

A new issue appears in the higher order case. The p + and the p" are the next larger and the next
smaller integer!to the non-integer p of the true period pT, and of course the two values differ by

! !
one, p + = p" +1 . Going back two periods, there will again be an integer just before and just after
the value 2p, but these need not be equal to 2 p + and 2 p" , and will be denoted by (2 p) + and
(2 p) " , with (2 p) + = (2 p) " +1. The repetitive control law in the z-domain becomes
! + + + +
U(z) = ["1 (a1z # p +1
+ b1z # p ) +!" 2 (a 2 z #(2 p )
!
+1
+ b2 z #(2 p ) )]U(z)
!
+ + + +
(28)
! ! + $z["1 (a1z # p +1
+ b1z # p ) + " 2 (a2 z #(2 p ) +1
+ b2 z #(2 p ) )]E(z)

and the RC transfer function with a zero-phase low-pass filter introduced, the difference equation,
and the characteristic polynomial governing stability become
!
+ +
U(z) F(z)z"(z)[#1z (2 p ) $ p (a1z + b1 ) + # 2 (a2 z + b2 )]
R(z) = = + + +
(29)
E(z) z (2 p ) $ F(z)[#1z (2 p ) $ p (a1z + b1 ) + # 2 (a 2 z + b2 )]
+ +
"p+
{z (2 p ) " F(z)[#1z (2 p ) (a1z + b1 ) + # 2 (a 2 z + b2 )][1" z$(z)G(z)]}E(z)
+ +
(30)
"p+
= {z (2 p ) " F(z)[#1z (2 p ) (a1z + b1 ) + # 2 (a 2 z + b2 )]}[YD (z) "V (z)]
! + +
"p+
z (2 p ) " F(z)[#1z (2 p ) (a1z + b1 ) + # 2 (a 2 z + b2 )][1" z$(z)G(z)] = 0 (31)

! If the new issue did not arise, then


p
one could separate the terms from the higher order and
from the interpolations as follows ("1z + " 2 )(a1z + b1 ) = H 2 (z)I 2 (z) , and then taking the absolute
!
value for z on the unit circle would produce a maximum value over all frequencies of unity, and
one could conclude that interpolation would not interfere with stability of second order repetitive
control. The higher order factor and the interpolation factor appear in similar places in the
! properties. However, the interpolation coefficients are a function of
equations, and have similar
one parameter, the distance of the true period from an integer period, whereas, the higher order
RC law is more general because it contains multiple coefficients that can be chosen
independently. Reference 12 suggests that instead of using a pure average of data from multiple
periods back, i.e. using equal weights for each period, one can use unequal weights to optimize
for different effects. One can use weights that decay as one goes back in repetitions in order to
decrease and flatten the amplification of disturbances at unaddressed frequencies between the
harmonics of the addressed frequencies. And one can make use of negative weights to decrease
the sensitivity to precise knowledge of the true period. Figure 4 shows the amplification and
attenuation of the forcing function for all frequencies, addressed and unaddressed, for first order
RC. Figure 5 is the corresponding plot for second order RC with equal weights, and one sees that
the maximum amplification is reduced from 2 to about 1.6. Figure 6 shows third order RC with
equal weights, which further reduces the maximum amplification. Also shown is the use of
decaying weights to make the maximum amplification smaller and more even. And finally the use
of negative weights is shown to widen the notches to make the response less sensitive to precise
knowledge of the period. These plots relate to the way the form of the forcing function eliminates,
or attenuates, or amplifies the amplitude of the disturbance. The final effect of this attenuation or
amplification is determined by using the sensitivity transfer function.

The sufficient stability condition for the second order repetitive control with linear
interpolation is given as
+
#p+
F(z)["1z (2 p ) (a1z + b1 ) + " 2 (a 2 z + b2 )][1# z$(z)G(z)] <1 (32)

!
Suppose that the weights "1 and " 2 are chosen to be positive. Then the new term can be bounded
as follows, for z on the unit circle as follows

+
#p+
! ! "1z (2 p ) (a1z + b1 ) + " 2 (a 2 z + b2 )
+
#p+
$ "1z (2 p ) (a1z + b1 ) + " 2 (a 2 z + b2 ) (33)

= "1 a1z + b1 + " 2 a 2 z + b2 $ "1 + " 2 =1

For this case, we can make the following conclusion: Suppose a compensator and a zero-phase
low-pass filter are designed to satisfy Eq. (8). Then, if one decides to use a second order RC with
positive weights, and!one decides to use linear interpolation, then Eq. (32) is satisfied. And it will
become clear that satisfying (32) is a sufficient condition for stability. Note that (33) can easily be
generalized to cubic interpolation as well, based on the numerical demonstration that Eq. (25)
holds. Of course there will be steady state error after convergence, resulting from the approximate
nature of interpolation creating a nonzero forcing function in the difference equation.

RELATIONSHIP OF THE MONOTONIC DECAY CONDITION TO THE TRUE


STABILITY BOUNDARY IN RC WITH INTERPOLATION

Reference 7 demonstrated that the heuristic monotonic decay condition (8) is a sufficient
condition for stability for any integer period p, and this is generalized here to include
interpolation in the repetitive control law in order to address non integer periods. A logical
approach to determining an if and only if condition for asymptotic stability is to directly apply the
Nyquist stability criterion for digital control to the feedback loop in Figure 1, i.e. apply the
criterion to R(z)G(z) = "1 . Because of the p poles on the unit circle and the need to go around
each of these with the Nyquist contour, this approach is impractical. Reference 7 presents a
different approach that avoids this difficulty and presents a proof that satisfying the heuristic
monotonic decay condition Eq. (8) does in fact ensure asymptotic stability even though it does
not!necessarily establish monotonic decay unless the quasi-static assumption is satisfied. Divide
+
the characteristic polynomial in (21), including interpolation and low pass filtering, by z p to
obtain
+
P(z) "1# z # p F(z)I j (z)[1# z$(z)G(z)] = 0 ! (34)

Once this is put over a common denominator, the numerator is the characteristic polynomial of
the repetitive control system. The contour needed to enclose all parts of the z-plane outside the
! +1 around the upper half of the unit circle, out to infinity on a branch cut
unit circle, goes from
along the negative real axis, then clockwise around at infinity, back in along the branch cut, and
around the lower half of the unit circle. We assume that the roots of the denominators of
F(z), "(z) and G(z) are all inside the unit circle, and furthermore assume that the degree of the
denominator of P(z) is greater than that of the numerator. These assumptions are normally
satisfied: if G(z) is a feedback control system, it should have all poles inside the unit circle, it
would also be normal to pick the compensator to be stable, and in addition, p is usually a large
! !
number. Then the contour at infinity is mapped onto the origin, and the branch cut goes back and
! the real axis from P(z) = P("1) (corresponding to Nyquist frequency) to the origin.
forth along
!
By the principle of the argument, the change in the phase angle of P(z) as z goes around the

!
!
contour will then equal the number of roots of the numerator inside the contour, i.e. the number of
unstable roots of the characteristic polynomial of the repetitive control system. If there is no
encirclement of the origin, then the repetitive control system is asymptotically stable. Now
rewrite the P(z) of Eq. (34) as

P(z) =1" P* (z)


+ (35)
! P* (z) = z " p F(z)I j (z)[1" z#(z)G(z)]

The second form represents a general stability test to use on repetitive control systems:
!
Since the denominator is of higher order than the numerator, one only needs to consider
the part of the contour associated with going around the unit circle, i.e. the frequency
response. Make a polar plot of the frequency response of P *(z) . If it does not encircle
the point +1, then the system is asymptotically stable. If it does encircle +1, the system is
unstable. This supplies an if and only if condition for asymptotic stability.

One can make a sufficient condition by asking ! that the plot of P *(z) be less than one in
magnitude at all frequencies up to Nyquist, in which case it cannot possibly encircle the point +1.
This produces the sufficient condition for asymptotic stability

#p+ !
P* (e i"T ) = e i"T F(e i"T )I j (e i"T )[1# z$(e i"T )G(e i"T )]
(36)
i"T i"T i"T i"T
= F(e )I j (e )[1# z$(e )G(e )] <1

Furthermore, if this condition is violated for some frequency range, it is very likely to create
+
instability, because the phase of the term z " p for the normally large values of p + moves very
!
fast, spreading the part of the curve outside the unit circle dramatically. See Ref. 11 for examples
of this effect.
! ! asymptotic stability of
Equation (36) establishes that Eq. (22) is a sufficient condition for
a first order RC system with both interpolation and zero-phase low-pass filtering. And, the
previous sections showed that if Eq. (8) is satisfied for all frequencies up to Nyquist, then so is
(36) when using linear or cubic interpolation. In addition, using a parallel development, one can
establish that Eq. (32) is a sufficient condition for asymptotic stability, and then satisfying (8) is a
sufficient condition for stability of second order RC with positive weights, including linear or
cubic interpolation.

EFFECT OF INTERPOLATION ON STEADY STATE ERROR

Although one has stability when one uses interpolation, the difference equation whose
solution is the tracking error will have a non-zero forcing function. One can predict how much
error the interpolation produces as a function of the addressed frequency, once steady state
response is reached, by using the sensitivity transfer function
+
z p " I j (z)
E(z) = + [YD (z) "V (z)] (37)
z p " I j (z)[1" z#(z)G(z)]

!
+
z p " I j (z)
S(z) = +
(38)
z p " I j (z)[1" z#(z)G(z)]

We examine this error for the fundamental frequency and for all harmonics for several systems.
- Base Case 1 refers to a system in which one is able to completely cancel the system
! a compensator "(z) . The effect is to have z"(z)G(z) =1 .
dynamics by using
- Base Case 2 applies to a third order system in continuous time having no zeros, and fed
by a zero order hold. This produces a discrete time transfer function with one zero outside
the unit circle located at -3.732 asymptotically as T tends to zero. All zeros and poles
inside the unit circle are ! !
cancelled with the compensator. The phase of the zero outside is
canceled using the method in Ref. 6, which introduces a zero inside the unit circle at the
reciprocal location, and introduces a pole at the origin.
- Base Case 3 considers a third order system again and again cancels everything except the
zero outside the unit circle. Instead of canceling just the phase of this zero, an FIR filter
with eight coefficients is designed that approximates the inverse of the frequency
response of the uncancelable zero. This is done using the method of Ref. 8.
- And, finally, a first order system a /(s + a) is used with a = 8.8 , which produces stable
RC without needing a compensator.
Figures 7 through 10 give the magnitude of the sensitivity transfer function using linear and cubic
interpolation, for each of these systems, as a function of the addressed frequency. If the
! !
interpolation did a perfect job, these plots would be zero at all frequencies. One can see how
much error there will be in steady state, introduced by the need to perform interpolation. The
sample rate is 100 Hz so that Nyquist frequency is 50 Hz. The addressed frequency is varied from
0.01 Hz ( p =10000 ) to 50 Hz ( p = 2 ). For a period of n time steps, for n an integer, the
associated frequency is 100/n, and these frequencies are zero in Figure 7 for Base Case 1. The
peaks between these frequencies reach 0.9 for linear interpolation and somewhat less for cubic,
indicating that interpolation is not accurate enough to have RC significantly attenuate the error at
! frequencies. Note that
these ! at lower frequencies, cubic interpolation is significantly better than
linear interpolation.

Figure 8 is the corresponding figure for Base Case 2. The high frequency peaks are
approximately the same, but lower frequencies look somewhat better. Note, however, that Base
Case 2 only cancels phase and does not adjust magnitude, so some of the improvement comes
from using a slower learning rate for some frequencies. Figure 9 uses the more precise design of
Base Case 3 that uses an FIR filter that approximately inverts the frequency response of the
system, with the learning rate being equalized at all frequencies. The maximum peak is further
improved, and in addition the peaks are made slightly narrower. All of the above figures attempt
to cancel the dynamics of the system. Some RC design methods simply use whatever
compensator is needed to stabilize the process, and do not attempt to cancel the system. Figure 10
examines RC on a first order system, which is stable without any compensator. In this case, RC
with interpolation fails to significantly attenuate the error for many frequencies. One can
conclude that if one has to use interpolation because the period is not an integer number of
sample times, one will get much better performance using a design that aims to cancel the system
dynamics.
THE INFLUENCE OF INTERPOLATION ON THE WATERBED EFFECT

The above studies the steady state error at the addressed frequencies. Now consider what
happens when there is some extra frequency present in the disturbance, other than the frequencies
having the addressed period, i.e. the fundamental and harmonics. The Bode integral theorem
applies to essentially all real time feedback control systems, and it states that if errors are
attenuated for some frequencies, for example at the addressed frequencies, then errors at some
other frequencies must be amplified, Ref. 17. This is often called the waterbed effect. Using the
sensitivity transfer function, one can evaluate how the use of interpolation influences the nature
of the amplification of other frequencies in the error. Here we consider a period of 8.3T
corresponding to 12.048 Hz at a sample rate of 100 Hz. There are three harmonics, the last one
being very close to Nyquist frequency. If no interpolation were required, then the minimum on
each of the valleys would be zero. They are not zero, and how far they are from zero is
determined in the previous section. The valleys are deeper when cubic interpolation is use, and as
the frequency goes up the valleys get shallower, as demonstrated in Figure 11 for Base Case 1,
Figure 12 for Base Case 2, and Figure 13 for Base Case 3. The waterbed effect governs the
amplification at other frequencies, and for Base Case 1 the amplification is below 1.2, but it
approaches 2 in both Base Cases 2 and 3. Figure 14 examines the first order system without using
a compensator. It appears that the need for interpolation has severely affected the depth of the
notches making much poorer performance attenuating the first and higher harmonics. Again, one
concludes that if one must use interpolation, it is important to use a compensator that tries to
invert the system in order to get good attenuation at addressed frequencies. On the other hand,
this narrowing of the notches allows the performance at other frequencies to stay much closer to
the ideal value of unity than in the Base Cases, i.e. staying much closer to not amplifying errors at
unaddressed frequencies.

CONCLUSIONS

The usual approach to repetitive control assumes that the period of the disturbance to be
addressed is an integer number of sample time intervals. There are effective ways to design such
repetitive controllers that aim to satisfy the sufficient stability criterion Eq. (8), making use of a
compensator that can be design using the method in Ref. 8 based on optimizing this criterion, and
then using a zero-phase low-pass filter designed using the method of Ref. 9 in order to robustify
the design to parasitic poles or residual modes. There is no reason to expect the disturbance
period to be an integer number of sample times, and one would need to use the nearest integer to
apply the above methods. This paper considers use of linear or cubic interpolation in order to
improve upon this nearest integer method. Several conclusions are made regarding the
performance of repetitive control using interpolation:

- If one uses Eq. (8) without regard to the possible need to use interpolation, and one
creates a compensator design and a low pass filter design that satisfies Eq. (8), then
introducing linear interpolation or cubic interpolation into the control law will not
destabilize the system.
- If one designs a first order RC system satisfying Eq. (8), and then modifies it to a second
order RC design using positive weights that sum to unity, then the second order RC
design is also stable. And if one then introduces linear or cubic interpolation, the RC
system will also be asymptotically stable.
- Violating the stability condition Eq. (8) when the period is an integer number of sample
times, p, is very likely to cause instability because of the fast phase change of the term z p

!
in the Nyquist style stability analysis (Ref. 11). When one uses linear or cubic
interpolation, the equivalent condition that is similarly close to the stability boundary is
given by Eq. (22), which can be easier to satisfy because the interpolation acts like an
additional low pass filter. This can allow one to use a higher cutoff frequency in the zero
phase cutoff filter design.
- However, in most applications one would not want to take advantage of this opportunity.
It requires that the period stay fixed at the same value between p + and p" . If the period
moves to one end or the other, then the stability condition reduces to Eq. (8), and the
difference between the two conditions goes away. A design that satisfies Eq. (22) but not
Eq. (8) is not robust to changes in period.
- When the disturbance is a integer number of time !steps, the ! error at the addressed
frequencies, fundamental and harmonics, is zero if one does not use a zero-phase low-
pass filter cutoff of the learning, and is zero for all frequencies below the cutoff when the
filter is a perfect cutoff filter. When interpolation is introduced because the error is not an
integer number of steps, the error at the addressed frequencies is degraded. Cubic
interpolation has less degradation than linear. At high frequencies approaching Nyquist,
the degradation can be very substantial. Note, however, that one might want a cutoff
somewhere around one half Nyquist anyway, for robustness to parasitic poles, to not
work the hardware too hard, and because zero error at sample times when using a zero
order hold to cancel the error, is not very effective at getting zero error between samples
at high frequencies.
- Many RC designs create a compensator that aims to cancel the dynamics of the system,
while other designs just make a compensator that has the property that it produces a
stable learning process, or use no compensator at all when the RC is stable without
compensation. Numerical comparisons suggest that the performance at the frequencies
being addressed by the RC design is much better overall when the cancellation type
designs are used. An example that is stable without a compensator has very narrow
notches in the frequency response at the addressed frequencies, making the notch depth
sensitive to the impreciseness of interpolation. However, the response to disturbances at
unaddressed frequencies was much flatter with less amplification for most frequencies.
- A competitor to the type of repetitive control treated here, is matched basis function
repetitive control (Ref. 18). This alternative aims to find the components of the error at
the frequencies being addressed, and then injects a corrective signal of the right
amplitude and phase to cancel the disturbance, taking into account the change in
amplitude and phase going through the system. Since one has identified the sinusoid in
the disturbance, one has the ability to sample it at any time of interest, i.e. interpolation is
replaced by simply sampling the known continuous time function at any time of interest.
This benefit over having to interpolate is obtained at some expense. One must treat each
frequency separately, rather than treat the fundamental and harmonics simultaneously.
There are inaccuracies in attempting to find the frequency components of interest. And
there is additional complexity because the governing equations have time varying,
periodic, coefficients.

In conclusion, when one needs to use interpolation in a repetitive control design, a recommended
approach is to aim to satisfy the sufficient condition for stability, Eq. (8), without regard to
interpolation. Then introducing interpolation as needed, will not destabilize the system. Using
cubic interpolation will give better performance than using linear interpolation, but one must
expect somewhat degraded performance by comparison to situations in which there is no need for
interpolation.
REFERENCES

1. T. Inoue, M. Nakano, and S. Iwai, “High Accuracy Control of a Proton Synchrotron Magnet
Power Supply,” Proceedings of the. 8th World Congress of IFAC, Vol. 20, 1981, pp. 216-
221.
2. R. H. Middleton, G. C. Goodwin, and R. W. Longman, “A Method for Improving the
Dynamic Accuracy of a Robot Performing a Repetitive Task,” University of Newcastle,
Newcastle, Australia, Dept. of Electrical Engineering TR EE8546, 1985. Also, 1989,
International Journal of Robotics Research, Vol. 8, pp. 67-74.
3. S. Hara, and Y. Yamamoto, “Synthesis of Repetitive Control Systems and its Applications,”
Proc. 24th IEEE Conference on Decision and Control, 1985, pp. 326-327.
4. S. Hara, T. Omata, and M. Nakano, “Stability of Repetitive Control Systems,” Proceedings of
the. 24th IEEE Conference on Decision and Control, 1985, pp. 1387-1392.
5. M. Tomizuka, T.-C. Tsao, and K. K. Chew, “Analysis and Synthesis of Discrete time
Repetitive Controllers,” Journal of Dynamic Systems, Measurement, and Control, Vol. 111,
1989, pp. 353-358.
6. K. Chew and M. Tomizuka, “Digital Control of Repetitive Errors in Disk-Drive Systems,”
IEEE Control Systems Magazine, Vol. 10, 1990, pp. 1620
7. R. W. Longman, “Iterative Learning Control and Repetitive Control for Engineering
Practice,” International Journal of Control, Special Issue on Iterative Learning Control, Vol.
73, No. 10, July 2000, pp. 930-954.
8. B. Panomruttanarug and R. W. Longman, “Repetitive Controller Design Using Optimization
in Frequency Domain,” AIAA/AAS Astrodynamics Specialist Conference and Exhibit,
Providence, Rhode Island, Aug. 16-19, 2004.
9. B. Panomruttanarug, and R. W. Longman, “Frequency Based Optimal Design of FIR Zero-
Phase Filters and Compensators for Robust Repetitive Control,” AAS/AIAA Astrodynamics
Specialist Conference, Paper Number AAS 05-263, August 2005.
10. H.-P. Wen and R. W. Longman, “Repetitive Control Methods When the Disturbance Period
is not an Integer Multiple of the Sample Time,” Advances in the Astronautical Sciences, Vol.
108, 2001.
11. S. Songschon and R. W. Longman, "Comparison of the Stability Boundary and the
Frequency Response Stability Condition in Learning and Repetitive Control," International
Journal of Applied Mathematics and Computer Science, Vol. 13, No. 2, 2003, pp. 169-177.
12. M. Steinbuch, "Repetitive Control for Systems with Uncertain Period-Time," Automatica,
Vol. 38, No. 12, 2002, pp. 2103-2109.
13. S. J. Oh and R. W. Longman, “Analysis of Stability and Performance in Higher Order
Repetitive Control,” Advances in the Astronautical Sciences, Vol. 110, 2003.
14. M. Steinbuch, S Weiland, J. van den Erenbeemt and T. Singh, “On Noise and Period Time
Sensitivity in High Order Repetitive Control,” Proceedings of 43rd IEEE Conference on
Decision and Control, December, 2004.
15. C.-P. Lo and R. W. Longman, "Root Locus Analysis of Higher Order Repetitive Control,"
Advances in the Astronautical Sciences, Vol. 120, 2005, pp. 2021-2040.
16. C.-P. Lo and R. W. Longman, “Frequency Response Analysis of Higher Order Repetitive
Control,” Advances in the Astronautical Sciences, this volume.
17. T. Songchon and R. W. Longman, "On the Waterbed Effect in Repetitive Control using Zero-
Phase Filtering," Advances in the Astronautical Sciences, Vol 108, 2002, pp. 1321-1340.
18. M. Nagashima and R. W. Longman, "Stability and Performance Analysis of Matched Basis
Function Repetitive Control in the Frequency Domain," Advances in the Astronautical
Sciences, Vol. 119, 2005, pp. 1581-1600.
V (z )

YD (z ) E (z ) U (z ) + Y (z )
+ +
R (z ) G (z )
-

Figure 1 Basic block diagram of a typical repetitive controller.

1
a1=0.0

0.8
a1=0.1
Magnitude

0.6
a1=0.2

0.4
a1=0.3

0.2
a1=0.4

a1=0.5
0
0 0.5 1 1.5 2 2.5 3 3.5
Normalized frequency θ (rad)

Figure 2 Plot of I 2 ( z ) vs. normalized frequency θ .

1
α =0.0

0.8
α =0.1
Magnitude

0.6
α =0.2

0.4
α =0.3

0.2
α =0.4

α =0.5
0
0 0.5 1 1.5 2 2.5 3 3.5
Normalized frequency θ (rad)

Figure 3 Plot of I 4 ( z ) vs. normalized frequency θ .


2 1.6

1.8
1.4

1.6
1.2
1.4
1
1.2
Magnitude

Magnitude
1 0.8

0.8
0.6
0.6
0.4
0.4

0.2
0.2

0 0
0 0.5 1 1.5 2 2.5 3 3.5 0 0.5 1 1.5 2 2.5 3 3.5
Normalized frequency θ (rad) Normalized frequency θ (rad)

Figure 4 Plot of z − 1 , p = 4 , z = exp(iθ ) − H 2 ( z) , p = 4 ,


p 2p
Figure 5 Plot of z
vs. normalized frequency θ . z = exp(iθ ) vs. normalized frequency θ .
1.5 1.4

1.2

1
1

0.8
Magnitude

Magnitude
0.6

0.5
0.4

0.2

0 0
0 0.5 1 1.5 2 2.5 3 3.5 0 0.5 1 1.5 2 2.5 3 3.5
Normalized frequency θ (rad) Normalized frequency θ (rad)

5
Magnitude

0
0 0.5 1 1.5 2 2.5 3 3.5
Normalized frequency θ (rad)

Figure 6 Plot of z
3p
− H 3 ( z ) , p = 4 , z = exp(iθ ) vs. normalized frequency θ ;
α 1 = α 2 = α 3 = 1 / 3 (above left), α 1 = 3 / 6, α 2 = 2 / 6, α 3 = 1 / 6 (above right),
α 1 = 2.93, α 2 = −2.93, α 3 = 1 (below).
0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
Magnitude

Magnitude
0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25 30 35 40 45 50
Frequency (Hz) Frequency (Hz)

Figure 7 Steady state error due to interpolation vs. addressed frequency, Base Case 1 with linear
interpolation (left), cubic interpolation (right).

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
Magnitude

Magnitude
0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25 30 35 40 45 50
Frequency (Hz) Frequency (Hz)

Figure 8 Steady state error due to interpolation vs. addressed frequency, Base Case 2 with linear
interpolation (left), cubic interpolation (right).
0.8 0.7

0.7 0.6

0.6
0.5

0.5
0.4
Magnitude

Magnitude

0.4
0.3
0.3

0.2
0.2

0.1 0.1

0 0
0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25 30 35 40 45 50
Frequency (Hz) Frequency (Hz)

Figure 9 Steady state error due to interpolation vs. addressed frequency, Base Case 3 with linear
interpolation (left), cubic interpolation (right).
1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
Magnitude

Magnitude
0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25 30 35 40 45 50
Frequency (Hz) Frequency (Hz)

Figure 10 Steady state error due to interpolation vs. addressed frequency, G ( s ) = 8.8 /( s + 8.8) ,
T = 1 / 100 , repetitive control law (7), φ = 1 . Linear interpolation (left), cubic interpolation (right).

1.4 1.4

1.2 1.2

1 1

0.8 0.8
Magnitude

0.6 Magnitude 0.6

0.4 0.4

0.2 0.2

0 0
0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25 30 35 40 45 50
Frequency (Hz) Frequency (Hz)

Figure 11 Response to disturbances at all frequencies, addressed period 8.3T, Base Case 1 with
linear interpolation (left), cubic interpolation (right).
2 2

1.8 1.8

1.6 1.6

1.4 1.4

1.2 1.2
Magnitude

Magnitude

1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25 30 35 40 45 50
Frequency (Hz) Frequency (Hz)

Figure 12 Response to disturbances at all frequencies, addressed period 8.3T, Base Case 2 with linear
interpolation (left), cubic interpolation (right).
2 2.5

1.8

1.6 2

1.4

1.2 1.5
Magnitude

Magnitude
1

0.8 1

0.6

0.4 0.5

0.2

0 0
0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25 30 35 40 45 50
Frequency (Hz) Frequency (Hz)

Figure 13 Response to disturbances at all frequencies, addressed period 8.3T, Base Case 3 with linear
interpolation (left), cubic interpolation (right).

1.4 1.8

1.6
1.2

1.4
1
1.2

0.8
Magnitude

Magnitude
1

0.6 0.8

0.6
0.4
0.4

0.2
0.2

0 0
0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25 30 35 40 45 50
Frequency (Hz) Frequency (Hz)

Figure 14 Response to disturbances at all frequencies, addressed period 8.3T, G ( s ) = 8.8 /( s + 8.8) ,
T = 1 / 100 , repetitive control law (7), φ = 1 . Linear interpolation (left), cubic interpolation (right).

You might also like