Download as pdf or txt
Download as pdf or txt
You are on page 1of 20

AAS 05-230

ROOT LOCUS ANALYSIS OF HIGHER


ORDER REPETITIVE CONTROL

Chun-Ping Lo † and Richard W. Longman *

Repetitive control (RC) is an effective method of eliminating the influence of periodic


disturbances to a feedback control system. Spacecraft often have vibrations from
imbalance in the moving parts of a cryogenic pump, or slight imbalance in a momentum
wheel, and RC is one candidate for producing active vibration isolation of fine pointing
equipment. RC normally uses error data from the previous period of the disturbance to
make corrections in the present period, whereas higher order RC considers data from
more than one previous period. Although the idea of higher order RC has appeared
multiple times in the literature, no general assessment of the potential of such designs has
appeared. This paper first presents the mathematical formulation of higher order RC, and
then studies second and third order RC by root locus methods, developing the pole-zero
patterns that can be generated by adjusting the weights on previous data. A companion
paper performs a parallel analysis using frequency response methods. It is proved that a
previously developed compensator design strategy for first order RC based on root locus
departure angle information, can be used without modification for higher order RC. The
use of negative weights on data from one of the previous cycles has been suggested in the
literature to have advantages in robustness to inaccuracy or fluctuations in the period of
the disturbance. The mechanism for this is clearly seen from a root locus approach, but
simple examples suggest that this insensitivity comes in exchange for more difficulty in
producing a stable design.

INTRODUCTION

The field of repetitive control (RC) aims to eliminate the influence of a periodic disturbance of
known period on the output of a feedback control system. Real time adjustments are made to the command
to the feedback control system aiming to converge to the command needed to cancel the effect of the
disturbance. Spacecraft often have one main source of vibrations such as imbalance in the moving parts of
a cryogenic pump, or slight imbalance in a momentum wheel. The resulting vibrations compromise the
performance of fine pointing equipment on board. Repetitive control is one main approach among several
that can be applied to this class of problems; other approaches include Multiple Error LMS, disturbance
identification, and the clear-box algorithm [1,2]. To apply such methods on spacecraft, one can mount the
fine pointing equipment on a six degree-of-freedom platform such as the Ultra Quiet Platform, a Stewart
Platform used in the experiments in [1,2]. Then one has the task of designing a controller for each of the six
actuators that can learn to move in such a way that the platform is stationary while the spacecraft on which
it is mounting in vibrating.
Typical RC adjusts the command to a control system based on the error observed one period back
in the disturbance. Various publications have suggested using data from multiple previous periods when
making updates to the command each time step, and this is referred to as higher order RC [3-7]. Although
the concept of higher order RC has appeared multiple times in the literature, it is not simple to determine
what all of the possible behaviors for all possible choices of weights might be, and no clear evaluation of
the possible advantages and disadvantages of higher order RC has been established. This paper contributes
to the understanding by examining the range of possible root locus configurations for first, second, and


Post-doctoral researcher, Department of Mechanical Engineering, Columbia University, New York, NY 10027.
*
Professor of Mechanical Engineering, also Professor of Civil Engineering and Engineering Mechanics, Columbia University, New
York, NY 10027
third order RC. A companion paper [8] addresses the same issue using frequency response analysis. Taken
together, these papers address: the possible pole and zero locations, the nature of the root locus plot,
stability criteria including stability with negative gains and stability robustness, convergence rate, final
error level in the presence of noise, and the waterbed effect.
Most often RC systems are designed based on a discrete-time model of the system, and the design
aims to cancel the system dynamics as much as possible. In [9-11] compensator design methods were
developed based on root locus departure angle analysis, and methods were generated to make the design by
applying a sequence of RC laws to the real world, and applying appropriate adjustments to the RC law. No
system model was needed. This paper proves that the same design method applies to higher order RC. It is
concluded that: 1. Higher order RC can be viewed as a first order RC with a high order compensator having
the same number of poles and zeros distributed on circles of different radii. 2. These extra poles and zeros
have no influence on the departure angle of poles on the unit circle. The locations of these extra poles and
zeros are determined by the choice of weighting factors and are summarized for second and third order RC.
A geometric proof that the net contribution of the extra poles and zeros to departure angles is zero in higher
order RC is provided.
References [6,7] consider higher order RC in terms of a weighted average on data from more than
one previous period. The weighted average concept implies that the weighting factors for different periods
should all be positive. Reference [7] provides a summary of performance in terms of convergence rate,
final error in the presence of noise, and the waterbed effect for positive weights. Allowing negative
weighting factors, this paper re-evaluates the performance of higher order RC. A normalization concept is
introduced that improves one’s ability to make meaningful comparisons, and as a result [8] re-evaluates
some conclusions in [7]. Empirical numerical evidence suggests that higher order RC does not have an
advantage over first order RC in improving the convergence rate or the final error level. However, with an
appropriate choice of negative weighting factors, higher order RC can suppress the amplitude of
unaddressed frequencies in the lower frequency range and reduce the sensitivity to fluctuations in the
period of the disturbance being addressed.

MATHEMATICAL FORMULATION OF HIGHER 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 by the repetitive controller, G(z) is the
closed loop transfer function of the feedback control system, V(z) is any deterministic periodic disturbance
and Y(z) is the output. A periodic disturbance might enter somewhere within the feedback control loop, but
there is always an equivalent periodic output disturbance which is used here.
Before going to higher order RC, consider first order repetitive control formulations. The most
basic form makes use of the concept of the discrete time equivalent of a separate integral control operating
on each time step of the disturbance period. Suppose that the period of the disturbance (or command, or
both when both are present) is pT , where T is the sample time interval, and p is the number of time steps
per period. Then the simplest form of RC is

u( k ) = u( k − p) + φe( k − p + 1) ; R( z ) = zφ /( z p − 1) (1)

The command at time step k is equal to the command one period back, plus a repetitive control gain φ
times the error one time step ahead of one period back (the T dependence in the arguments is dropped for
notational simplicity). The one step ahead compensates for the usual one step delay going through G( z ) .
Note that stability of the repetitive control system is determined by the characteristic polynomial
R( z )G( z ) = −1 . And the root locus produced by varying the repetitive control gain φ has the zeros of the
original feedback control system and in addition it has a new zero at the origin. The poles are the poles of
the original feedback control system, which are assumed to be inside the unit circle, and in addition there
are p poles on the unit circle introduced by the repetitive controller. In practice, it is convenient to design
the RC law in the frequency domain, using the form
z pU ( z ) = F( z )[U ( z ) + Φ( z ) E( z )] ; R( z ) = F( z )Φ( z ) /( z p − F( z )) (2)

This generalizes equation (1) by replacing φz with Φ( z ) that can include not only a gain, but also a
compensator. Methods of designing the compensator include those in [10-13]. 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 this is done at the expense of not trying to learn to
eliminate periodic errors above the cutoff. From Fig. 1, one can write [1 + G( z ) R( z )]E( z ) = YD ( z ) − V ( z ) ,
and using equation (2) one obtains the difference equation whose solution is the tracking error of the
control system

{z p − F( z )[1 − Φ( z )G( z )]}E( z ) = [ z p − F( z )][YD ( z ) − V ( z )] (3)

When there is no filter, i.e. F( z ) = 1 as in equation (1), then the forcing function on the right side of (3)
becomes zero because it is the difference of the value of periodic functions at the present time and shifted
one period ahead. Hence, in this case the error converges to zero provided the characteristic polynomial in
curly brackets on the left has all roots inside the unit circle.
Higher order RC considers the error not only one period back, but also includes the corresponding
errors at multiple periods back. The form for Nth order RC corresponding to (1) is
N
u( k ) = ∑ α [u(k − jp) + φe(k − jp + 1)]
j =1
j (4)

where N is the number of periods one wishes to include, and the α 1 , α 2 ,K,α N are weights to be chosen by
the designer. One may think of this as creating a weighted average, in which case each weight should be
non-negative. However, as pointed out in [5], there is no need to restrict the weights to be positive. The
generalizations of equations (2) and (3) to include a compensator and zero-phase low-pass filter are

F( z )Φ( z )[α 1 z ( N −1) p + α 2 z ( N −2 ) p + L + α N ]


R( z ) = (5)
z Np − F( z )[α 1 z ( N −1) p + α 2 z ( N −2 ) p + L + α N ]
{z Np − F( z )[1 − Φ( z )G( z )][α 1 z ( N −1) p + α 2 z ( N −2 ) p + L + α N ]}E( z )
(6)
= {z Np − F( z )[α 1 z ( N −1) p + α 2 z ( N −2 ) p + L + α N ]}[YD ( z ) − V ( z )]

and the characteristic equation R( z )G( z ) = −1 is again given by setting the curly bracket on the left to zero.
The conditions needed to converge to zero error again require that all roots of the characteristic polynomial
lie inside the unit circle, and that the forcing function on the right of difference equation (6) is zero. The
latter occurs when no cutoff filter is used, so that F( z ) = 1 , and in addition this time it is necessary to
restrict the choice of the α j to satisfy

α1 + α 2 + L + α N = 1 (7)

HIGHER ORDER RC AS 1st ORDER WITH COMPENSATOR, AND NORMALIZATION

Consider RC of the form of equation (4) with a compensator Φ( z ) , and F ( z ) = 1 , and consider
the range of possible second order and third order RC laws. Then the repetitive controllers for first, second,
and third order RC from equations (1,5) become

Φ( z )
R1 ( z ) = (8)
zp −1
Φ( z )(α 1 z + α 2 ) Φ( z )(α 1 z p + α 2 )
p
R2 ( z ) = 2 p = p = R1 ( z )C2 ( z ) (9)
z − α1z p − α 2 ( z − 1)( z p + α 2 )
Φ( z )(α 1 z 2 p + α 2 z p + α 3 ) Φ( z )(α 1 z 2 p + α 2 z p + α 3 )
R3 ( z ) = = = R1 ( z )C3 ( z ) (10)
z 3 p − α1z 2 p − α 2 z p − α 3 ( z p − 1)[ z 2 p + (1 − α 1 )z p + α 3 ]

It is the constraint (7) that allows us to factor the denominator as in the second forms of equations (9,10).
And this factored form allows one to consider higher order RC as first order RC with a high order
compensator, C2 ( z ) or C3 ( z ) (however, when a low pass filter F is introduced one can no longer make
such an interpretation). The higher order RC root locus will always have the zero at the origin and the p
poles on the unit circle of the first order RC, but in addition it will have p extra poles and p extra zeros in
the case of second order, or 2p extra poles and 2p extra zeros in the case of third order RC.
To facilitate the process of comparing RC laws of different orders, it is helpful to introduce a
normalization to the repetitive control gain. The compensators above do not have a DC gain of unity so that
the un-normalized gain φ for different order RC corresponds to different learning rates for different orders
for zero frequency. We define a normalized gain φ N to adjust the DC level for all RC orders to be the
same, and in addition to adjust the DC gain to unity when applied to the feedback control system G( z ) :

φ = φ N K1 K2 (11)
−1
 C (1) = 1 + α 2 , 2nd order RC
K1 =  −1 2
(12)
C3 (1) = 2 − α 1 + α 3 = 1 + α 2 + 2α 3 , 3rd order RC
K2 = 1 /[Φ(1)G(1)] (13)

In [8], numerical experiments using this normalization produce essentially identical learning rates for
various different order RC laws with various weights.

BASE CASES

There are many methods to design the compensator Φ( z ) to use in first order RC, but much of the
literature aims to create a compensator that is equal to 1 / G( z ) . If this inverse is stable, then one can
actually use this approach, and when there is one or more zero of G( z ) outside the unit circle, then one
must compromise.

Base Case No. 1: When one can make a stable compensator that is the system inverse, the root locus plot
for first order RC is defined by φ /( z p − 1) = −1 , which has p poles on the unit circle and no zeros, and as
the gain is raised, all roots go radially inward reaching the origin when φ = 1, and then continue radially
outward, all leaving the unit circle when φ reaches 2. When p is odd, they simply continue along their
original path after reaching the origin, and when p is even they proceed radially outward along radial lines
halfway between the radial lines from the p original poles of z p = 1 to the origin. This is called the base
case in [9], and will be called Base Case No. 1 here. Writing the associated characteristic polynomials for
first, second, and third order RC
zp −1+ φ = 0
z 2 p − α 1 z p − α 2 + φ (α 1 z p + α 2 ) = 0 (14)
z 3p
− α1z 2p
− α 2 z − α 3 + φ (α 1 z
p 2p
+ α2z + α3 ) = 0
p

we can see how this generalizes to higher order RC. When the unnormalized gain φ is set to unity, all of
these RC systems have all roots of the characteristic equation located at the origin. The paths used to get
there can be rather different as shown below, but in each case one can have all roots at the origin at the
same time. And this corresponds to learning as fast as possible, i.e. in a deadbeat manner, for the
characteristic equation above. Normally it is roots that start from the p poles on the unit circle that have the
slowest decay and determine the learning speed. But one must remember that the design does pole-zero
cancellation of the system G( z ) making the transients of the system uncontrollable. Therefore, the actual
learning rate is limited to the speed of decay of these transients.
Base Case No. 2: For the majority of systems, the inverse 1 / G( z ) is unstable and cannot be used as a
compensator. When a continuous time system is fed by a zero order hold, one can create a discrete time
model that predicts the output at the sample times without any approximation. This model will generically
have a number of zeros that is one less zero than the number of poles (producing a one time step delay from
change in input to change in output). Thus, with a pole excess of pe in continuous time, there will be
pe − 1 zeros introduced in the disretization. Reference [14] predicts where these zeros are located as the
sample time T tends to zero, which is at the zero locations of the discrete time equivalent of the transfer
function 1 / s pe fed by a zero order hold. All of these zeros are introduced on the negative real axis. With a
pole excess of 1, no zero is introduced. With a pole excess of two, one zero is introduced asymptotically at
-1. With a pole excess of three, two zeros are introduced, asymptotically one is inside the unit circle at -
0.2679 and the other is at -3.7321. Note that these two locations are reciprocals of each other. With a pole
excess of four, one zero is introduced inside, one outside at the reciprocal location, and one at -1
asymptotically. For purposes of illustration in this paper we consider a pole excess of three. Reference [13]
designs the Φ( z ) to cancel all poles and zeros that can be cancelled in G( z ) . Any zero outside the unit
circle that cannot be cancelled, has its influence on the phase cancelled, so that the product Φ( z )G( z ) has
zero phase. The gain of the compensator can then be adjusted to give the product a unity DC gain. For a
zero factor ( z + c) of G( z ) outside the unit circle, a factor ( z −1 + c) is introduced in the numerator of Φ( z ) ,
and then the product normalized at DC. This product can be written as

( z + c)( z −1 + c) = c( z + c)( z + 1 / c) / z (15)

The product is equal to [1 + ( z + z −1 )c + c 2 ] = 1 + 2(cos ωT )c + c 2 for z = exp(iωT ) on the unit circle, which
is real and positive, and hence the term in Φ( z ) succeeds in canceling the phase although not the
magnitude influence of the zero outside the unit circle in G( z ) . Reference [11] also shows that this
compensator design succeeds in canceling any influence of the zero outside the unit circle on the departure
angles of the p poles on the unit circle. We define Base Case No. 2 to correspond to designing a
compensator by this approach. Consider a pole excess of three (e.g. a third order system in continuous time
with no zeros), and consider the limit as T tends to zero. The zero outside the unit circle would normally be
compensated according to (15) by introducing a zero inside at the reciprocal location, and introducing a
pole at the origin. Since asymptotically, there is already a zero at the reciprocal location, one would not
cancel it, and then all that is needed is to introduce the pole at the origin. Assuming perfect cancellation of
everything else in G( z ) , then what remains is the expression (15) with c = 3.7321 , together with any
normalization. To summarize, Base Case No. 2 for a pole excess of 3 corresponds to the product Φ( z )G( z )
given by (15) and normalized to unity at DC.

Base Case No. 3: Reference [12] develops a compensator design method in the form of an FIR filter
together with a chosen number of poles at the origin. When an appropriate number of gains are allowed, the
design method will produce the same pattern of a zero outside the unit circle on the negative real axis being
compensated by a zero inside at the reciprocal location, and a pole at the origin, but this pattern is not just
done on the negative real axis, but it is repeated at evenly spaced angle intervals around the origin. The
more times it is repeated, the better the product Φ( z )G( z ) approximates the frequency response of a unity
transfer function. We define Base Case No. 3 to correspond to having a chosen number of equations (15)
rotated to evenly spaced angles around 360 degrees. For purposes of this paper, we again consider what
Base Case 3 produces for a pole excess of 3, and we choose to repeat the pattern every 90 degrees. Base
Case No. 3 specialized in this way has 4 poles at the origin, and 8 zeros at radii equal to c = 3.7321 , and
1 / c = 0.2679 , and angles of 0, ±90 , and 180 degrees.

POLE AND ZERO LOCATIONS FOR SECOND ORDER RC

From (9) for second order RC it is clear that there are still the p poles on the unit circle as before
(which we will call the original p poles), and in addition there are p additional poles introduced as well as p
additional zeros introduced. The defining equations are
Constraint: α1 + α 2 = 1
Extra Zeros: z p = −α 2 / α 1 (16)
Extra Poles: z p = −α 2

The case of α1 = 0 is eliminated as being of no interest since it is like a first order RC except that it fails to
use the most recent period's data. Also, α 1 = 1 is not allowed, since then the higher order RC reduces to
first order RC. Thus, neither α 1 nor α 2 can be equal to one, and neither can be equal to zero. In addition,
we require both here and for third order RC below, that all poles of R( z ) be inside the unit circle, with the
exception of the original p poles on the unit circle. Then, at least for sufficiently small repetitive control
gains φ , one will have a stable repetitive control process if the root locus departs from the original p poles
with a radially inward component. Therefore, it is also required that α 2 < 1 . Combining these results, the
range of possible choices of the gains α 1 , α 2 is restricted to

If α 2 > 0 , then 0 < α 1 < 1 and 0 < α 2 < 1


If α 2 < 0 , then 1 < α 1 ≤ 2 and −1 ≤ α 2 < 0 (17)

The zeros equation in (16) can be rewritten as z p equals α 2 / α 1 exp(i(π ± 2 lπ )) for α 2 > 0 , and
α 2 / α 1 exp( ±i 2 lπ ) for α 2 < 0 , l = 0, 1, 2, 3,... Then the locations of the zeros and the poles are given by

 α 2 / α 1 exp(inθ ) for α 2 < 0


1/ p
 α 2 exp(inθ ) for α 2 < 0
1/ p

z= ; z =  1/ p (18)
 α 2 / α 1 exp(i(nθ + π / p)) for α 2 > 0  α 2 exp(i(nθ + π / p)) for α 2 > 0
1/ p

respectively, where θ = 2π / p , and n = 0, 1, 2,..., p − 1 . From these equations we can generate the following
pole zero patterns. Example root locus plots associated with these patterns are given in Fig. 2, and will be
discussed in the next section.

(i) When α 2 is chosen greater than zero:


- All introduced zeros and poles lie on the radial lines half way between those going from the origin to
the original p poles on the unit circle
- When α 2 / α 1 > α 2 the radius to the zeros is larger than that to the poles
- When α 2 / α 1 > 1 , the introduced zeros are outside the unit circle (Fig. 2a)
When α 2 / α 1 < 1 , the introduced zeros are inside the unit circle (Fig. 2b)
When α 2 / α 1 = 1, the introduced zeros are on the unit circle (Fig. 2c)
(ii) When α 2 is chosen less than zero
- All extra poles and zeros introduced by going to second order RC lie on the same radial lines as the
original poles on the unit circle
- The introduced zeros are inside the unit circle and are on a circle of radius smaller than that of the
introduced poles ( α 2 < 0 implies that α 1 > α 2 so the zeros are inside the poles which are inside the
unit circle), Fig. 2d.
- The limiting case α 2 = −1 repeats the original poles on the unit circle. The other gain is forced to be
α 1 = 2 , which puts the introduced zeros on a circle whose radius is the pth root of α 2 / α 1 = 1 / 2 .
When p gets large these zeros get close to the poles on the unit circle.

ROOT LOCUS PLOTS FOR SECOND ORDER RC

Base Case No. 1: In Base Case No. 1, all poles and zeros of G( z ) have been cancelled by Φ( z ) and the
root locus is defined by equations (14) for changing repetitive control gain φ . For first order RC, the root
locus has been described above, and all p poles on the unit circle go radially inward to the origin. Figure 2
gives root locus plots for second order RC. For clarity of illustration, p is set to 4, but one should keep in
mind that p, the number of time steps in a period, could be a very large number. As usual, poles are
indicated by × in the plot, although sometimes one of the crossed lines coincides with the departure
direction making it hard to recognize. As indicated in the figure, Fig. 2a has weights that increase going
back in time, 2b has weights that decay, 2c has equal weights that corresponds to taking a pure average of
data from two previous cycles, and 2d is a case of using negative weights as suggested in [5]. In the first
three the introduced poles are closer to the origin than the introduced zeros, and vice versa for the last of
the figures. The departure angles from all poles on the unit circle are radially inward, from which one can
conclude that all designs are stable for small enough positive gain φ . Of course we already know that when
φ = 1 all roots will be at the origin.
Consider the case of negative weights. If one chose α 1 = 2 and α 2 = −1, then the characteristic
polynomial when φ = 0 becomes z 2 p − 2 z p + 1 = ( z p − 1) 2 , and the original p poles on the unit circle are
being repeated. The departure angles will be tangential to the unit circle in this case. As the gain is
increased, the curvature of the root locus is smaller than the unit circle and the roots will go inside the unit
circle. By picking α 1 = 1.85 and α 2 = −0.85 , the introduced pole is pulled some distance radially inward,
and Fig. 2d shows the pole on the unit circle departing radially in, while the pole inside departs radially
outward to meet it. The learning rate is related to the slowest decaying root of the characteristic equation,
and the decay in each time step is equal to the magnitude of this root. Both of these cases will have slow
learning for small gains. How high a gain one can use depends on how successful the model cancellation is.
Negative weights are suggested in order to widen the notches around the frequencies of period p time steps,
the fundamental and all harmonics to Nyquist frequency. For first order RC with no filter F, the forcing
function in equation (3) has transfer function z p − 1 . Consider the associated frequency transfer function by
setting z = exp(iωT ) , and consider frequencies in the neighborhood of any notch, i.e. ωTp = l2π + ε where
l is an integer and ε is a measure of the deviation from the notch frequency. Expanding in a Taylor series
through linear terms in the neighborhood of this frequency, z p − 1 becomes equal to ε . This produces a
cusp to zero in the magnitude plot for steady state error frequency response of equation (3). Now consider
second order RC in the case when the poles on the unit circle are repeated. The case with the introduced
poles brought inward a small distance will have somewhat similar properties. With F = 1 , α 1 = 2 ,
and α 2 = −1, the right hand side of equation (6) has transfer function ( z p − 1) 2 , and expanding this around
the notched frequencies produces ε creating a quadratic minimum. At least for small enough deviations,
2

the quadratic minimum will be more tolerant to inaccuracy in the notch position of the design, or to
variations in the actual frequency of the disturbance. How effective this is in improving the notch width is
shown in the companion paper [8].

Base Cases Nos. 2 and 3: Figure 3 studies the root locus for first order RC in Base Cases No. 2 and No. 3
(root locus plots for third order RC will be given later). These apply when one is unable to make a
compensator that cancels the system dynamics. Figure 3a is for Base Case No. 2 with Fig. 3b giving a
detailed view, and Fig. 3c is for Base Case No. 3 with Fig. 3d giving a detailed view. Note that the root
locus for Base Case No. 3 needs to be a negative gain root locus, because the coefficient of the highest
power of z in the numerator is negative. In each case, p was set to 6 in order to allow Base Case No. 3 to be
causal. Neither of these cases can get all the roots to the origin simultaneously as in Base Case No. 1. Base
Case No. 3 can come very close to inverting the frequency response of the system, but this is not the same
in inverting poles and zeros. A sampling was done to determine the fastest possible learning rate for each
case. For Base Case No. 2 the gain was incremented in intervals of 0.01 and the largest magnitude of the
roots recorded for each gain. The fastest learning rate occurred for a gain of 0.26 which produced a decay
by a factor of 0.8892 each time step. For Base Case No. 3, the gain was incremented in intervals of 0.0001
and at a gain of 0.0052 the minimized maximum radius was 0.6012. This is substantially faster learning
than for Base Case No. 2.

First Order System: In first order RC when applied to a first order system, there is no need for a
compensator and no need for a low pass filter as far as the mathematics is concerned. The learning law (1)
will converge to zero error. Figure 4 examines how this generalizes to higher order RC. Consider a
continuous time system G( s) = a /( s + a) with a = 8.8 , fed by a zero order hold running at 100 Hz sample
rate. Figure 4a presents the root locus plot for first order RC, Fig. 4b is for second order RC corresponding
to Fig. 2a with weights that grow going back in time, and Fig. 4c corresponds to the negative weight case of
Fig. 2d. The root loci associated with Fig. 2b and 2c are very similar to Fig. 4b but with the introduced
poles and zeros brought inward the appropriate distances.

Third Order System: Many design methods for first order RC produce stability of the process, but do not
try to invert the system. Consider a third order system

G( s) = [a /(s + a)][ω o2 /(s 2 + 2ζω o s + ω o2 )] (19)

with a = 8.8, ω o = 37, ζ = 0.5 , fed by a zero order hold sampling at 100 Hz. The method in reference [10]
developed a compensator for this system, and it is shown below that compensators developed by that
method apply to higher order RC as well. The compensator has 2 poles at the origin, two poles at ±45
degrees and radius 0.8, complex zeros at ±3.6 degrees and radius 0.9, and three real zeros at 0.2, 0.4, and
0.5. Again, Fig. 4d is for first order RC on this system with this compensator, Fig. 4e is for second order
RC with the weights of Fig. 2a, and Fig. 4f is for the negative weight case of Fig. 2d. In Fig. 4f, the roots
starting at the poles on the imaginary axis move inward to a radius of about 0.994 and then move toward
instability. It appears that it is harder to stabilize the learning process when using these negative weights,
and one should aim to use an inversion based design as in the base cases above. For the positive weights of
Fig. 4e there is a stable range of repetitive control gains, rather similar to that of first order RC.

POLE AND ZERO LOCATIONS FOR THIRD ORDER RC

The analog of equation (16) for third order RC are

Constraint: α1 + α 2 + α 3 = 1 (20)
Extra Zeros: z 2 p + (α 2 / α 1 )z p + (α 3 / α 1 ) = 0 (21)
Extra Poles: z 2 p + (1 − α 1 )z p + α 3 = 0 (22)

Note that we assume that α 1 ≠ 0 , since making it zero implies that one ignores the data from the most
recent period, which does not make physical sense. The form of the equations for both the zeros and the
poles is quadratic in x = z p

(α 2 / α 1 , α 3 / α 1 ) for zeros


x 2 + ax + b = 0 ; ( a, b) =  (23)
 (1 − α 1 , α 3 ) for poles

And the solutions for x and their magnitudes will be denoted by

(
x s = − a + a 2 − 4b / 2 ) ; (
x d = − a − a 2 − 4b / 2 ) (24)
Ms = x s ; Md = x d

The solutions are divided into 6 cases developed below.

Case 1. Solutions x s , x d form a complex conjugate pair. This corresponds to

a 2 − 4 b < 0 ; b > 0 ; Ms = Md = b (25)

Then the p poles or zeros come from each of the following equations
ϕ = atan2 ( 4b − a 2 , − a ) (26)
z=b 1 /( 2 p )
exp[i(nθ + ϕ / p)] ; z = b exp[i(nθ − ϕ / p)]
1 /( 2 p )
(27)
n = 0, 1, 2, 3,..., p − 1 : θ = 2π / p

Case 2. Solutions x s = x d are repeated which corresponds to

a 2 − 4 b = 0 ; x s = x d = − a / 2 ; Ms = Md = a / 2 (28)

and the corresponding zeros or poles can be obtained by substituting into the following equations which
also apply to the distinct real root cases below.

 M 1 / p exp(inθ ) for x s > 0  M 1 / p exp(inθ ) for x d > 0


z =  1/ p s ; z =  1/ p d (29)
 Ms exp[i(nθ + π / p)] for x s < 0  Md exp[i(nθ + π / p)] for x d < 0

Again, p roots for the poles or zeros come from each equation, using n = 0, 1, 2, 3,..., p − 1 .
The following four cases apply to distinct real roots for x. Note that neither x s nor x d can be zero
since this would require that b = 0 and hence α 3 = 0 , which reduces to second order RC.

Case 3. Distinct real roots xs < 0, x d < 0 which requires a 2 − 4b > 0, a > 0, b > 0 (30)
Case 4. Distinct real roots xs > 0, x d > 0 which requires a 2 − 4b > 0, a < 0, b > 0 (31)
Case 5. Distinct real roots xs > 0, x d < 0, Ms < Md which requires a 2 − 4b > 0, a > 0, b < 0 (32)
Case 6. Distinct real roots xs > 0, x d < 0, Ms > Md which requires a 2 − 4b > 0, a < 0, b < 0 (33)

Note that x s < 0, x d > 0 results in a contradiction. The above cases apply both to zeros and to poles, and it
is possible for a particular choice of gains α 1 , α 2 , α 3 , that the zeros fall into one case and the poles into
another. These cases are illustrated in Fig. 5 which is shown for poles, including the p poles on the unit
circle. The pattern of the additional 2p poles are shown, for all cases but Case 2 which is a limiting case,
and the same patterns apply for the 2p additional zeros.
In the following, we examine the ranges of gains that can apply to each case. These ranges are
necessary conditions to be in a particular case, but are not expected to be sufficient conditions. To develop
the ranges, the inequalities in the definitions of each case are used, as well as the constraint equation (20).
For poles, we further require that none of the introduced poles be outside the unit circle, or equivalently, we
require that Ms ≤ 1, Md ≤ 1 . The bounds used as necessary condition are different for different cases:

Complex Roots: b ≤1 (34)


Real Roots with b > 0 : a ≤ 1, b > 0 (35)
Real Roots with b < 0 : a ≤ 1, a 2 < 4(1 + b) , b < 0 (36)

The first of these is immediate from the magnitude of the roots as given by (25). For the second of these
equations, both roots have the same sign, and the maximum value they can assume is given by a . The
third case is more complicated. When b < 0 the two roots have opposite signs with the square root term
larger in magnitude than the first term on the right side of the equal signs in (24). Then we can ask for
a / 2 ≤ 1 / 2 and ( a 2 − 4b)1 / 2 / 2 ≤ 1 as necessary conditions. These produce the conditions given in (36).
The following subsections apply the inequalities summarized in this paragraph to both zeros and poles for
each of the six cases and develop necessary conditions on the ranges of values of α 1 , α 2 , α 3 for the zeros or
poles to be in each case.
Case 1
Zeros: The second condition in (25) requires that α 1 and α 3 have the same sign, and the first
condition becomes α 22 < 4α 1α 3 . Both sets of zeros have the same radius. This case is unusual in the sense
that all other cases have the introduced zeros (or poles) either lying on the radial lines from the origin going
to the original poles, or on radial lines that are halfway between those to the original poles. In this case, p of
the introduced zeros (or poles) lie on radial lines with angles shifted in a counterclockwise direction by
angle ϕ / p , and p of the introduced zeros (or poles) have the angles shifted clockwise by this same
amount. This is illustrated for the case of poles in Fig. 5.
Poles: For poles to be in this case and be in or on the unit circle, the weights need to satisfy

(1 − α 1 ) 2 < 4α 3 ; 0 < α 3 ≤ 1 ; 0 ≤ α 1 + α 2 < 1 ; − 1 < α 1 < 3 ; − 3 < α 2 < 1 (37)

The first of these equations is the first of equations (25). The second uses the first of (25) for the left
inequality and (34) for the right. The third uses the constraint (20) in the second of (37). The fourth uses the
maximum value α 3 = 1 from the second, on the right hand inequality of the first, to produce a range on α 1 .
The last inequality in (37) solves (20) for α 3 and substitutes into the first of (37) to obtain
(1 + α 1 ) 2 < 4(1 − α 2 ) from which we conclude that α 2 < 1 . Isolating α 2 in the third of (37) and using the
fourth of (37) produces the lower bound in the last of (37).

Case 2
Zeros: One can pick the value of α 2 / α 1 , then the first equation in (28) gives the ratio α 3 / α 1
which must be positive, meaning that α 1 and α 3 must have the same sign. Then the constraint (20)
determines the value of α1 and hence all α j . The initial choice of α 2 / α 1 immediately determines the
values of x s = x d = −(α 2 / α 1 ) / 2 , their magnitudes Ms = Md , and the pth root of these magnitudes, which
gives the radius of the circle on which the zeros lie. When α 1 and α 2 are chosen with opposite signs,
x s = x d > 0 , and the roots are given by the upper part of (29), meaning that the zeros can be placed
anywhere along the radial lines from the origin to the original poles on the unit circle. When they are
chosen with the same signs, they can be placed anywhere along radial lines with angles half way between
those to the original poles. Note that one should not pick α 2 / α 1 = 0 since this makes α 2 = α 3 = 0 , and the
problem degenerates to first order RC. The value α 2 / α 1 = 2 puts the zeros on the unit circle, and the value
α 2 / α 1 = −2 is at a singularity.
Poles: One can pick the value of α 1 , and then x s = x d = (α 1 − 1) / 2 which determines the radius
of the circle for the new poles. One needs −1 ≤ α 1 ≤ 3 for these poles to be inside or on the unit circle. The
value of α 3 is determined by the first of equations (28), and (20) then determines the remaining α 2 . The
roots are at [(1 − α 1 ) / 2]1 / p exp[i(nθ + π / p)] when −1 ≤ α 1 < 1 , and lie on the radii to the original poles
(see Case 4 in Fig. 5 with the introduced poles at the same radius). The roots are at [(α 1 − 1) / 2]1 / p exp[inθ ]
when 1 < α 1 < 3 and lie on radial lines halfway between those of the original poles (Case 3 in Fig. 5 with
the introduced poles on the same radius). When α 1 equals -1 or +3 the roots are on the unit circle, when it
is zero one gets the case where the most recent period of data is ignored, and when it is equal to +1 the RC
law degenerates into a first order RC. Note that there are actually two more special cases, or subcases: let
α 1 = 3, α 2 = −3, α 3 = 1 and one has Case 4 with all poles on the unit circle, and letting
α 1 = −1, α 2 = 1, α 3 = 1 sends the introduced poles of Case 3 so they are on top of each other and on the
unit circle.

Case 3
Zeros: The last two inequalities of (30) imply that all three α j have the same sign, and (20) then
says that 0 < α j < 1 , which corresponds to the weighted average concept for higher order RC. In addition
one needs that α 22 > 4α 1α 3 from the real root condition in (30). There are p new zeros on an inner circle of
radius Ms1/ p , and p new zeros on an outer circle of radius Md1/ p , and one can have these circles either inside
or outside the unit circle. Both radii are necessarily less than (α 2 / α 1 )1 / p . The zeros for each circle are
along radial lines halfway between the radii to the original poles on the unit circle (see Fig. 5).
Poles: The weights for this case must satisfy

(1 − α 1 ) 2 > 4α 3 ; 0 < α 1 < 1 ; 0 < α 3 < 1 / 4 ; 3 / 4 < α 1 + α 2 < 1 ; −1 / 4 < α 2 < 1 ; Md ≤ 1 (38)

The first comes from the third from the last of the inequalities in (30), the second comes from (35) and
a > 0 , the third from b > 0 and the first of (38) with α 1 at its lower limit, the fourth from the third with
α 3 solved from constraint (20), and the fifth uses the second in the fourth.

Case 4
Zeros: The last two inequalities of (31) imply that α 1 and α 3 have the same sign and α 2 has the
opposite sign, and they must satisfy α 22 > 4α 1α 3 . The zeros are given by the upper equations in (29) and
hence both sets lie on the lines joining the origin to the original poles on the unit circle. The larger radius is
Ms and they both can be outside the unit circle (see Fig. 5).
Poles: The weights for this case must satisfy

(1 − α 1 ) 2 > 4α 3 ; 1 < α 1 < 2 ; 0 < α 3 < 1 / 4 ; 3 / 4 < α 1 + α 2 < 1 ; −1 / 4 < α 2 < 0 ; Ms ≤ 1 (39)

One side of the second of (39) comes from a < 0 and the other side from the limit on magnitude in (35).
Since b > 0 , α 3 > 0 and pure inequality holds. Other inequalities are obtained analogously to the previous
case.

Case 5
Zeros: This time α 1 and α 2 have the same sign and α 3 has the opposite sign. There are p new
zeros from the top left of equation (29) that are on the radial lines to the original poles, and at a larger
radius is another set of zeros from the bottom right of (29) which lie on radial lines that are halfway
between those to the original poles (see Fig. 5).
Poles: The weights for this case must satisfy

(1 − α 1 ) 2 > 4α 3 ; 0 < α 1 < 1 ; − 1 < α 3 < 0 ; 1 < α 1 + α 2 < 2 ; 0 < α 2 < 2 ; Md ≤ 1 (40)

The second of (40) uses a > 0 , and for the lower limit it uses the first of (36). The third uses the second and
third of (36) (see Fig. 5).

Case 6
Zeros: This time α 1 and α 3 have the same sign and α 2 has the opposite sign, and α 22 > 4α 1α 3 .
The new zeros are similar to the above case, with the difference that the new zeros on the radial lines to the
original poles on the unit circle are now at a larger radius than those on radial lines centered between those
to the original poles (see Fig. 5).
Poles: Following similar logic as before, the weights for this case must satisfy

(1 − α 1 ) 2 > 4α 3 ; 1 < α 1 < 2 ; − 1 < α 3 < 0 ; 1 < α 1 + α 2 < 2 ; − 1 < α 2 < 0 ; Ms ≤ 1 (41)

ROOT LOCUS PLOTS FOR THIRD ORDER RC

Figures 6, 7, and 8 give examples of root locus plots illustrating the cases in Fig. 5. Figure 6
applies to Base Case No. 1 designs. Figure 6a is for equal weights and both the zeros and the poles are Case
1 in Fig. 5. When the weights decay going back in repetitions, with α 1 , α 2 , α 3 being 3/6. 2/6, 1/6, Case 1
still applies for poles and zeros, and the plot is similar but pulled in a bit. With weights growing, 1/6, 2/6,
3/6, and when the weights have a dip in the middle 4/9, 1/9, 4/9, Case 1 still applies to both poles and zeros.
Figure 6b gives the negative weight design using 2.93, -2.93, 1 and again it is Case 1 for both
poles and zeros. If the weights are changed to 3, -3, 1, then the 2p extra poles introduced by going to third
order RC are on the unit circle, ahead and behind the original p poles. Suppose all three poles were on top
of each other, then one would depart radially inward, and the other two would depart with a radially
outward component at 60 degrees from the tangent to the circle, and hence the system would be unstable at
least for small gain. When the roots are separated a small distance from the original poles on the unit circle,
but still lie on the unit circle as in 3, -3, 1, one expects similar behavior. By pulling the weights in to 2.93,
−2.93 , 1 some tolerance is developed giving a small range of stability, but it still goes unstable for an
interval before becoming stable again. Pulling the poles in further could fix this. The negative weights do
have a benefit in widening the notches around the addressed frequencies. Ignoring stability issues, the
steady state frequency response with the extra poles on the unit circle will have zero magnitude at the
addressed frequency of the notch, and also at the two new neighboring frequencies, and when these
neighboring frequencies are close, the plot will stay small between the three frequencies. When the poles
are not on the unit circle, but pulled in somewhat, the frequency response will be similar but will not go all
the way to zero at the frequencies associated with the new poles. Hence, the approach can be effective in
widening the notches, but it aggravates the problem of producing stability.
In Figs. 6c, 6d, 6e, 6f weights with a bump in the middle, 1/6, 4/6, 1/6 produce Case 3, with a dip
in the middle 1.9, -1, 0.1 produce Case 4, weights 0.5, 1, -0.5 produce Case 5, and weights 1.2, -0.1, -0.1
produce Case 6. In all of these, the Case for the zeros is the same as the Case for the poles, and it is difficult
to create a situation where the two have different Cases. But it can be done, Fig. 6g with 1.7, -0.8, 0.1 is
Case 4 for the poles and Case 1 for the zeros.
Figure 7 examines all of the same example weights for Base Case No. 3 (p is changed to 6 instead
of 4 for causality, and recall that the root locus for this Case is a negative gain locus). The first plot gives
the overall view, which is similar for all Cases, and then Figs. 7b to 7h give zoomed in views for each of
the choices of weights. Figure 8 again examines all of the choices of weights, and applies them to the first
order system a /( s + a) without including a compensator. The negative weights of 2.93, -2.93, 1 that widen
the notches, again seems to have stability problems. Again it seems that aiming to widen the notches
aggravates the stability problem, and in order to use such weights one should have the best possible model
and aim to cancel it according to one of the Base Cases.

ROOT LOCUS DEPARTURE ANGLE ANALYSIS

The characteristic polynomial is usually of high degree because p is usually large, making it
difficult to solve for the roots and difficult to use typical classical control design methods. Reference [9]
suggests that although one may not be able to make a full root locus analysis, it is relatively easy to
examine the departure angles from the p poles on the unit circle. And if all departure angles have a
component going toward the interior of the unit circle, the RC law will be asymptotically stable for all
sufficiently small repetitive control gains φ . Reference [10] develops a library of compensators and their
influence on the departure angles, and shows how they can be used to fix instability in first order RC.
Reference [11] goes further and develops methods of using the compensator library without the benefit of a
model of the system. One runs a candidate repetitive control system, and as an instability appears one can
analyze the frequency content to determine which poles are departing from the unit circle with an outward
component. Several tests can then be used to know whether the departure angle needs to be increased or
decreased, and which compenator can be used to establish stability. It is the purpose of this section to
establish that the same compensator design strategy also applies to higher order RC. The procedure is to
design the compensator the same way one does for first order RC in [11]. Then one picks whatever higher
order RC configuration one might want in order to improve some performance characteristic, and the
combined design is then asymptotically stable for sufficiently small repetitive control gain φ .
This section concludes that: (1) Going to second or third order RC does not affect the departure
angles of the original poles on the unit circle. And, (2) any compensator design that produces stability for
all sufficiently small gains for first order RC also does so for second and third order RC, provided the
introduced poles are inside the unit circle. This applies to any design including ones based on [10,11].
Second order RC adds p extra poles evenly distributed on a circle inside the unit circle, and p extra
zeros on a circle either inside the unit circle or outside. Third order does the same, except p is replaced by
2p. There are three possible orientations of the extra sets of p poles and p zeros: (1) p roots can lie on radial
lines that are located halfway between the radial lines to the original poles on the unit circle (e.g., Fig. 5,
Case 3), (2) they can lie on radial lines going from the origin through the original poles on the unit circle
(e.g., Fig. 5, Case 4), or (3), they can be spaced angle ϕ ahead and angle ϕ behind the radial lines going to
the original poles on the unit circle (Fig. 5 Case 1).
First consider a set of p poles and p zeros inside the unit circle, and examine their total
contribution to the departure angle of any chosen pole on the unit circle. Use letter i as an index counting
the poles on the unit circle starting with i = 1 for the pole at +1, and progressing counterclockwise up to
Nyquist frequency at 180 degrees. Denote the angle made by the radial line from the origin to pole i as θ i .
Nearly all of the introduced poles (or zeros) inside the unit circle can be grouped in pairs that are symmetric
about the radial line to pole i. Figure 9a illustrates such a situation, with line db along the radial line, b
represents the pole on the unit circle, and a and c are the two poles (or zeros) from the inner circle. Lines
ae and cg are lines parallel to df starting at a and c. By the symmetry of the situation, lines ab and bc
have equal length, and together with line ac form an isosceles triangle. Line ac is then orthogonal to line
df . Then the angles α and β have the same magnitude, and the contribution to the root locus departure
angle from pole i produced by the pole or zero at a is θ i + α , and from the pole or zero at c is θ i − β
(where the contribution is either positive or negative depending on whether the roots are poles or zeros).
The net contribution from the pair of poles or zeros is 2θ i , so that one can choose to consider that each pole
or zero contributed angle θ i .
Consider that both the introduced poles and the introduced zeros follow the first case (1) described
above. Then all poles and zeros introduced can be put in pairs as in Fig. 9a, and the total contribution from
the zeros is pθ i , and from the poles is − pθ i , making a net zero contribution from all of the poles and zeros
introduced by going to higher order RC. As a result, the methods of [10,11] for designing compensators for
first order RC, can apply to higher order RC without modification in this case.
Now consider the case when the introduced poles or zeros are on the radial lines to the poles on
the unit circle. If p is even, then there will be two poles or zeros introduced on the radial line to the ith pole
on the unit circle, one on each side of the origin. It is obvious that each root has a θ i contribution, and
hence the above conclusion still applies. If p is odd as in Fig. 9b, there is one root on this line between the
origin and the ith pole, again contributing θ i , and again the above conclusion applies.
So far, we have considered the zeros introduced to be inside the unit circle. Now consider what
happens when p zeros are outside. For the moment, ignore the case that the zeros lie on the radial lines to
the poles on the unit circle. One can again draw a diagram with isosceles triangles for symmetric pairs of
zeros, as shown in Fig. 9c, and the arguments above again apply. Note that this is still true when the zeros
introduced are on the unit circle, and Fig. 9a still applies.
It remains to consider the case of zeros introduced outside the unit circle and on the radial lines
through the poles on the unit circle. If α 1 > 0 , then one needs a positive gain root locus, and the pole at +1
will depart radially outward, provided there is only one zero outside on the same radial line (i.e., the real
axis). From Fig. 9d it is clear that there is an extra 180 degree contribution in this case. This would produce
instability. However, for third order RC one can show that asking for α 1 > 0 , and x s > 1 , and the sum of
the weights equal to one, produces a contradiction, and that asking for x d > 1 implies that x s > 1 , and again
there is a contradiction. If α 1 < 0 and one must use a negative gain locus, one would have instability if
there were two roots outside the unit circle on the same radial line. However, asking for α 1 < 0 , x s > 1 , and
x d > 1 results in a contradiction. For second order RC, to be on the same radial line, z p = −α 2 / α 1 must be
greater than one, and the second condition in (17) must apply to match the sign, but it is impossible to make
it greater than one. The conclusion is that zeros outside the unit circle cannot be on the lines through the
poles on the unit circle, and the departure angles will not be changed by going to second or third order RC.

CONCLUSIONS

This paper presents an analysis of higher order RC using root locus plots, and the companion
paper [8] performs a parallel analysis using frequency response methods. It is shown here that higher order
RC can be viewed as first order RC with a high order compensator, provided that no zero phase filter is
used to cut off the learning at high frequencies. A normalization is introduced that facilitates comparison of
the performance of different order RC laws in [8]. Rather general results are presented for first order and
higher order RC designs when the compensator design aims to cancel the system dynamics. Three base
cases are defined, one for the case when the inverse of the system model is stable, and two that handle the
case when there are zeros outside the unit circle. The last of these is an improved approach to creating a
stable inverse. Reference [11] develops a different compensator design approach for first order RC that
does not require a system model, and instead is based on root locus departure angle information and
observing the evolution of frequency content of the error with time. It is proved here to apply also to higher
order RC. It has been suggested in the literature that negative weights in higher order RC can be used to
reduce sensitivity to error in one's knowledge of the period of the disturbance, or to fluctuations in the
period. Root locus analysis shows that the sensitivity can in fact be improved by appropriate choice of the
weights. Simple numerical examples suggest that producing this insensitivity is at the expense of increased
difficulty in producing stability.

REFERENCES

1. S. G. Edwards, B. N. Agrawal, M. Q. Phan, and R. W. Longman, "Disturbance Identification and


Rejection Experiments on an Ultra Quiet Platform," Advances in the Astronautical Sciences, Vol. 103,
1999, pp. 633-651.
2. H.-J. Chen, R. W. Longman, and B. N. Agrawal, "Approaches to Matched Basis Function Repetitive
Control," Advances in the Astronautical Sciences, Vol. 109, 2002, pp. 931-950.
3. T. Inoue, "Practical Repetitive Control System Design," Proceedings of the 29th IEEE Conference on
Decision and Control, Honolulu, Hawaii, 1990, pp. 1673-1678.
4. W. S. Chang, I. H. Such, T. W. Kim, "Analysis and Design of Two Types of Digital Repetitive Control
Systems," Automatica, Vol. 31, No. 5, 1988, pp.741-746.
5. M. Steinbuch, "Repetitive Control for Systems with Uncertain Period-Time," Automatica, Vol. 38, No.
12, 2002, pp. 2103-2109.
6. S. J. Oh and R. W. Longman, "Stability of Higher Order Repetitive Control," Modeling, Simulation and
Optimization of Complex Systems, H. G. Bock, E. Kostina, H. X. Phu, and R. Rannacher, Editors,
Springer Verlag, Heidelberg, Germany, 2005, pp. 383-392.
7. S. J. Oh and R. W. Longman, "Analysis of Stability and Performance in Higher Order Repetitive
Control," Advances in the Astronautical Sciences, Vol. 116, 2003, pp. 1291-1310.
8. C.-P. Lo and R. W. Longman, "Frequency Response Analysis of Higher Order Repetitive Control," to
appear.
9 . R. W. Longman and E. J. Solcz, "Small Gain Robustness Issues in the p-Integrator Repetitive
Controller," Proceedings of 1990 AIAA/AAS Astrodynamics Conference, Portland, OR, August 20-22,
1990, pp.537-551.
10. C.-P. Lo and R. W. Longman, "On the use of Root Locus Departure Angle Information to Design
Compensators in Repetitive Control," Advances in the Astronautical Sciences, Vol. 112, 2002, pp.
1383-1402.
11. C.-P. Lo and R. W. Longman, "Designing Root Locus Departure Angle Compensators in Repetitive
Control Based on Error Spectrum," Advances in the Astronautical Sciences, Vol. 119, 2005, pp. 1389-
1408.
12. B. Panomruttanarug and R. W. Longman, "Repetitive Controller Design Using Optimization in the
Frequency Domain," Proceedings of the 2004 AIAA/AAS Astrodynamics Conference, Providence, RI,
Aug. 2004.
13. M. Tomizuka, T. C. Tsao, K.-K. Chew, "Analysis and Synthesis of Discrete Time Repetitive
Controllers," Journal of Dynamic Systems, Measurement, and Control, Vol. 111, 1989, pp. 353-358.
14. K. Åström, P. Hagander, and J. Strenby, "Zeros of Sampled Systems," Proceedings of the Nineteenth
IEEE Conference on Decision and Control, 1980, pp. 1077-1081.
V(z)
YD(z) + E(z) U(z) + + Y(z)
R(z) G(z)
_

Fig. 1 Basic block diagram of a repetitive control system.

(a) (b)

(c) (d)

Fig. 2 Root locus plots of second order RC for Base Case No. 1.

(a) (b)

(c) (d)

Fig. 3 Root locus plots of first order RC for Base Case No. 2 and No. 3.
(a) (b)

(c) (d)

(e) (f)

Fig. 4 Root locus plots of 1st, 2nd order RC for a first order system
and a third order system with compensator.

θ/p

Case 1 Case 3 Case 4

Case 5 Case 6

Fig. 5 Pole and zero patterns for third order RC.


(a) (b)

(c) (d)

(e) (f)

(g)

Fig. 6 Root locus plots of third order RC for Base Case No. 1.
(a) (b)

(c) (d)

(e) (f)

(g) (h)

Fig. 7 Root locus plots of third order RC for Base Case No. 3.
(a) (b)

(c) (d)

(e) (f)

(g)

Fig. 8 Root locus plots of third order RC for the first order system.
g f

b θi
θi β
(a) (b)
c e
α
d
θi
a

c θi
f

α d e
π
θi
(c) α (d)
a
θi
b

Fig. 9 Illustration of departure angle contribution to poles on the


unit circle from the higher order roots.

You might also like