You are on page 1of 20

AAS 19-935

SECOND ORDER REPETITIVE CONTROL:


EVALUATION OF STABILITY BOUNDARY AND DEVELOPMENT
OF SUFFICIENT CONDITIONS

Ayman F. Ismail1, Richard W. Longman2, Peiling Cui3, Zhiyuan Liu4, and


Han Xu5

Spacecraft often have vibrations from slight imbalance in control moment gyros,
reaction wheels, or momentum wheels. Repetitive Control (RC) is an effective
method to eliminate the jitter produced, actively isolating fine pointing equipment
from spacecraft vibrations. The period of the disturbances is known since these
rotations are being commanded, and a first order RC adjusts the command of a
closed loop isolation system during the current period based on the error observed
at the corresponding times in the previous period, aiming to converge to zero
tracking error. The frequency response of first order RC has narrow notches at the
addressed frequencies having the given period, requiring accurate knowledge of
the disturbance period. With imperfect knowledge, or with fluctuations in the pe-
riod, the performance is compromised. Second order repetitive control is a design
that can reduce sensitivity to period fluctuations. A disadvantage is added stability
complications. This paper focuses on second order RC, initially including data
from two periods back. The approach developed here makes use of a partial frac-
tion expansion, developed by the third author and her research group, that allows
second order RC to use data only from one period back, with equivalent results as
if it were using data from two periods back. It also allows for parallel processing
if desired. Several contributions are made here to address the stability complica-
tions. An algorithm to evaluate the true stability boundary is presented. A general
sufficient stability condition from previous literature is reviewed for comparison
purposes. And a new sufficient stability condition is derived and then simplified
to make it much easier to use and independent of the number of time steps in a
period. The paper evaluates how close the sufficient conditions are to the true
stability boundary and confirms the effectiveness of using multiple sufficient sta-
bility conditions as a design tool for second order RC.

INTRODUCTION
Repetitive control (RC) is a relatively new field that aims to produce zero tracking error follow-
ing a periodic command, or in the presence of a periodic disturbance of known period. A major

1 Doctoral Candidate, Department of Mechanical Engineering, Columbia University, 500 West 120 th St., New York, New
York 10027, USA. ayman.ismail@columbia.edu
2 Professor of Mechanical Engineering, and Civil Engineering and Engineering Mechanics, Columbia University, 500

West 120th St., New York, New York 10027, USA. RWL4@columbia.edu
3 Associate Professor, School of Instrumentation and Optoelectronic Engineering, Beihang University, Beijing, China,

100191. cuiplhh@126.com
4 MS Candidate, Beihang University, Beijing, China, 100191, liuzhiyuan66@qq.com
5 MS Candidate, Beihang University, Beijing, China, 100191, 1067597037@qq.com

1
application in spacecraft is eliminating jitter, or vibrations caused by slight imbalance in control
moment gyros, or reaction wheels, or a momentum wheel. Experimental tests on a spacecraft float-
ing testbed were reported in Reference 1.
First order RC uses data from one period back to adjust the command in the current period,
aiming to make the error converge to zero asymptotically. One might think to use Nyquist stability
criterion to study stability and determine the range of parameters producing asymptotic stability
and convergence. However, a direct application of this criterion requires the use of a Nyquist con-
tour that not only goes around the unit circle in the z-domain, but also goes around 𝑝 poles that are
on the unit circle, where 𝑝 is the number of time steps in a period. This is impractical. A way to
reconfigure the analysis and circumvent the difficulty was developed in Reference 2. In addition,
it created a proof of a sufficient condition applicable for determining stability for all possible peri-
ods. It was shown in Reference 3, that this sufficient condition was very close to the necessary and
sufficient condition, the difference being of no importance in practical applications.
A difficulty in the use of RC to cancel periodic disturbances is that one must know the period
of the disturbance and it is assumed constant. In some applications, the period fluctuates with time.
In others, one has a sensor that reports the disturbance period, but it may not be reported with the
accuracy necessary to get good performance. These practical difficulties motivated Reference 4 to
study higher order RC, RC that uses data from more than one period back, and uses some negative
weight(s). From the perspective of magnitude frequency response of the sensitivity transfer func-
tion, the frequency amplitude goes to zero at the fundamental and all harmonics of the given period,
making a valley going to zero at the bottom for each frequency. In first order RC this valley can be
very narrow, meaning that small deviations in the period may move the response out of the valley
and even amplify instead of attenuate error. Higher order RC with properly chosen weights can
widen this valley and may significantly improve the error cancellation sensitivity to period error or
period fluctuations. A study of the design of higher order RC considering the tuning of gains for
this purpose was presented in References 5, 6, and 7. An alternative method to make RC insensitive
to period was compared in Reference 8.
This paper focuses on second order RC, using data from two periods back. The necessary and
sufficient stability condition is presented from Reference 6. A new sufficient condition is created
for second order RC paralleling the development of an effective sufficient condition for first order
RC systems in References 2 and 3. It makes use of a partial fraction expansion introduced in Ref-
erence 9 that allows second order RC to use only data from one period back, with equivalent results
as if it were using data from two periods back. This has an advantage in real time memory require-
ments and also opens possibilities for new simpler sufficient stability conditions that are more gen-
eral and less conservative. Here, this equivalent formulation is used while reformulating the
Nyquist stability condition in Reference 2 to handle this more complex situation. The formulation
also combines positive and negative weights into one stability condition that is simpler than existing
methods for higher order RC stability. The third author and co-workers, utilized the partial fraction
representation in Reference 9 to formulate a new, but different, sufficient stability condition that
explicitly defines the bounds for RC gains and phase angles to make a system stable. In a future
paper, Reference 10, this new sufficient condition will be compared to that introduced here, as well
as the other sufficient conditions discussed here.
Introducing second order RC complicates the stability boundary and stability robustness be-
cause the degree of the characteristic polynomial is doubled, now containing the power two times
the number of time steps in a period. This can be a large number, making the phase of such factors
spin many times going around the unit circle of the reformulated Nyquist contour. It is the purpose
of this paper to supply some evaluation of this issue. There are several general approaches to

2
address stability complications that are considered: (1) A computer algorithm to show how the true
stability boundary can be determined in spite of the added difficulties, (2) Evaluation of a previ-
ously developed sufficient stability condition for comparison purposes, (3) Development of a suf-
ficient condition for practical design purposes, (4) Simplification of developed sufficient condition
to make it easier to use and to eliminate dependence on the number of time steps in the period, (5)
Evaluation and comparison of the sufficient conditions to the true stability boundary demonstrating
effectiveness of the sufficient conditions for use in second order RC designs.

MATHEMATICAL FORMULATION OF SECOND ORDER REPETITIVE CONTROL


Consider the block diagram structure for a general RC system in Figure 1, where the repetitive
controller 𝑅(𝑧) adjusts the command 𝑢(𝑘𝑇𝑠 ), or 𝑈(𝑧) when transformed to the discrete-time do-
main, 𝑘 is an integer sample number and 𝑇𝑠 is the constant sample time interval. 𝐺(𝑧) represents
the closed loop transfer function of the feedback control system. The periodic disturbance 𝑉(𝑧) is
deterministic and represents the equivalent output disturbance to the periodic physical disturbance
that is usually originally between the feedback controller and the plant before they were combined
into an equivalent transfer function 𝐺(𝑧). Different scenarios can be addressed by RC design such
as the desired output 𝑌𝐷 (𝑧) can be periodic with an integer 𝑝 time steps per period and the disturb-
ance 𝑉(𝑧) is constant, the desired output can be constant and the disturbance is periodic with an
integer 𝑝 time steps per period, or both the desired output and disturbance are periodic with the
same integer 𝑝 time steps per period. A special treatment is needed for the case that 𝑝 is a non-
integer and will not be addressed in this paper.
In its simplest form, the repetitive control law for second order RC can be expressed as
𝑢(𝑘𝑇𝑠 ) = 𝛼1 𝑢((𝑘 − 𝑝)𝑇𝑠 ) + 𝛼2 𝑢((𝑘 − 2𝑝)𝑇𝑠 )
(1)
+ 𝜙[𝛼1 𝑒((𝑘 − 𝑝 + 𝛾)𝑇𝑠 ) + 𝛼2 𝑒((𝑘 − 2𝑝 + 𝛾)𝑇𝑠 )]
Thus the command given to the feedback system at the current time step can be computed as a
weighted sum of the command signal from one and two periods back plus a scalar RC gain 𝜙 times
the weighted sum of the error signal from one and two periods back (error is defined as the desired
output minus the actual output), but shifted forward by 𝛾 time steps representing the delay from
input to output of the feedback control system. Coefficients 𝛼1 and 𝛼2 can be thought of as produc-
ing a weighted average when they are both positive, but they can be positive and negative as intro-
duced in Reference 4 in order to widen notches, and must sum to one in order to converge to zero
error. It is common to design RC in the frequency domain and to generalize the RC gain in Equation
(1) from 𝜙𝑧 𝛾 to 𝜙𝐹(𝑧) that includes a constant gain and a compensator to adjust the phase and
amplitude to ensure stability. The compensator design that is used in this paper, consists of an FIR
(finite impulse response) filter based on the method described in Reference 11 that mimics the
inverse of the system frequency response inverse. A zero-phase low pass filter 𝐻(𝑧) can also be
introduced to cut off learning above a certain frequency to improve convergence of the system in
the presence of high frequency model error. By taking the z-transform of Equation (1), the general
form of the second order RC in the z-domain is expressed as
𝑧 𝑝 𝑈(𝑧) = 𝐻(𝑧)[𝑈(𝑧)(𝛼1 + 𝛼2 𝑧 −𝑝 ) + 𝜙𝐹(𝑧)𝐸(𝑧)(𝛼1 + 𝛼2 𝑧 −𝑝 )] (2)
The second order RC controller transfer function that relates the input error 𝐸(𝑧) to output 𝑈(𝑧)
in the repetitive controller block in Figure 2 is
𝑈(𝑧) 𝐻(𝑧)𝜙𝐹(𝑧)𝑄(𝑧) 𝐻(𝑧)𝜙𝐹(𝑧)(𝛼1 𝑧 𝑝 + 𝛼2 )
𝑅(𝑧) = = = 2𝑝
𝐸(𝑧) 1 − 𝐻(𝑧)𝑄(𝑧) 𝑧 − 𝐻(𝑧)(𝛼1 𝑧 𝑝 + 𝛼2 ) (3)
𝑄(𝑧) = 𝛼1 𝑧 −𝑝 + 𝛼2 𝑧 −2𝑝 , 𝛼1 + 𝛼2 = 1

3
where 𝑄(𝑧) is a second order RC gain function. The difference equation for the RC system can be
obtained from Figure 1 as (1 + 𝑅(𝑧)𝐺(𝑧))𝐸(𝑧) = 𝑌𝐷 (𝑧) − 𝑉(𝑧), and combined with Equation (3)
to produce
[𝑧 2𝑝 − 𝐻(𝑧)(𝛼1 𝑧 𝑝 + 𝛼2 )(1 − 𝜙𝐹(𝑧)𝐺(𝑧))]𝐸(𝑧)
(4)
= (𝑌𝐷 (𝑧) − 𝑉(𝑧))(𝑧 2𝑝 − 𝐻(𝑧)(𝛼1 𝑧 𝑝 + 𝛼2 ))
For a perfect low pass filter, 𝐻(𝑧) is one below the cutoff frequency with no phase change, and
zero above the cutoff frequency. Since the desired output and disturbance signals are periodic with
exact period 𝑝 time steps, the right side or forcing function in Equation (4) becomes zero below the
cutoff and remains unaffected above the cutoff. The characteristic polynomial can be written as
𝑧 2𝑝 − 𝐻(𝑧)(𝛼1 𝑧 𝑝 + 𝛼2 )(1 − 𝜙𝐹(𝑧)𝐺(𝑧)) = 0 (5)
The RC system will be asymptotically stable if the roots of the characteristic polynomial in Equa-
tion (5) lie inside the unit circle.
To gain insight into the RC transfer function and to highlight the improved sensitivity to period
fluctuations for higher order RC, the magnitude of the frequency response of the sensitivity transfer
function from command to error for second order RC in Equation (4) is compared to the corre-
sponding transfer function for first order RC. Note, the transfer function equation for first order RC
is not shown in this paper but can be easily derived. For providing a practical example, a third-
order transfer function model of a feedback system is considered as
𝑎 𝜔02
𝐺(𝑠) = ( )( 2 ) (6)
𝑠 + 𝑎 𝑠 + 2𝜁𝜔0 𝑠 + 𝜔02
Equation (6) describes the model of the command to response of the control system for each link
of a Robotics Research Corporation robot link as in Reference 2. The constants are defined as 𝑎 =
8.8, 𝜁 = 0.5 and 𝜔0 = 37.07. The continuous system model is fed by a zero-order hold at a sam-
pling rate of 100 Hz and is compensated by either a 12 gain or 8 gain FIR filter that approximates
the inverse of the steady state feedback control system frequency response. Figure 3 gives the mag-
nitude frequency response of the sensitivity transfer function from command to error for first and
second order RC designs using two different gains corresponding to 𝜙 = 1 and 𝜙 = 0.1 with 𝑝 =
10 time steps per period. The error is zero at the bottom of the notches corresponding to the funda-
mental frequency and all its harmonics up to Nyquist. For the same RC gains, the second order RC
system has a wider valley which makes the system less sensitive to period fluctuations. For the
chosen design parameters, the right-hand side of Equation (4) introduces |𝑧 𝑝 − 1|2 to the numerator
of the sensitivity transfer function which has a quadratic shape at the minimum. For a first order
RC, the equivalent sensitivity transfer function has |𝑧 𝑝 − 1| which shows as a cusp at the minimum.
Although the error magnitude is generally lower for first order RC between addressed frequencies,
both designs behave worse than having a feedback system alone. It is worth noting that one pays
significantly for the widened valleys. The influence of disturbances between frequencies addressed
by the RC design, can be amplified by a factor of 4. For low enough RC gains, i.e. below 𝜙 = 0.1,
one can design a system to have benefits of the feedback system between addressed frequencies
and the benefits of RC at addressed frequencies, provided one knows the period accurately. The
requirement for improved insensitivity to period error comes at the expense of both.
Four different stability conditions will be presented in the next sections. The first condition
represents the true stability boundary (if and only if condition) that guarantees asymptotic stability.
We develop a computer algorithm that evaluates this condition, necessarily sampling very fre-
quently, and adding up the phase changes from sample to sample. The second is a sufficient

4
condition that was developed via a rigorous mathematical approach in Reference 6 and is presented
in this paper for evaluation and comparison purposes. The third condition is also a new contribution
that parallels the development of a sufficient condition for first order RC in Reference 2. The fourth
is a simplification of the third condition to make it independent of period 𝑝 and thus easier to eval-
uate. Finally, an evaluation of all sufficient conditions and comparison to true stability boundary is
performed to demonstrate the effectiveness of sufficient conditions in second order RC designs.
FIRST STABILITY CONDITION
In this section, the true stability boundary (if and only if condition) that guarantees asymptotic
stability is developed using a computer algorithm as suggested in Reference 6. A logical approach
for determining the stability condition for the RC system, is to use the standard Nyquist stability
criterion for digital control applied to the feedback loop of Figure 1. Direct application of the
Nyquist criterion to the characteristic polynomial 𝑅(𝑧)𝐺(𝑧) = −1 for 𝑧 going around the Nyquist
contour is impractical, because it requires making a Nyquist contour that goes around all 𝑝 poles
on the unit circle as shown on the left of Figure 4. Instead, a reformulated problem is developed
below to circumvent this difficulty, following Reference 6. The characteristic polynomial from
Equation (5), is multiplied by 𝑧 −2𝑝 and rewritten as
𝑃(𝑧) = 1 − 𝑧 −2𝑝 𝐻(𝑧)(𝛼1 𝑧 𝑝 + 𝛼2 )(1 − 𝜙𝐹(𝑧)𝐺(𝑧)) = 0 (7)
When the terms in Equation (7) are put over a common denominator, the resulting numerator is the
characteristic polynomial for the RC system. The contour needed to enclose all parts of the z-plane
outside the unit circle is shown in the right plot of Figure 4, it goes from +1 around the upper half
of the unit circle, out to infinity on a branch cut along the negative real axis, then clockwise around
at infinity, back along the branch cut, and around the lower half of the unit circle. Note this time
there are no poles on the unit circle so that it is unnecessary to deal with the many small half circles.
As before, it is assumed that the roots of the denominators of 𝐻(𝑧), 𝐹(𝑧), and 𝐺(𝑧) are all inside
the unit circle. These conditions are satisfied if 𝐺(𝑧) is a feedback system with all poles inside the
unit circle and the compensator and zero-phase low pass filter were picked to be stable. The prin-
ciple of the argument states that change in the phase angle of 𝑃(𝑧) as 𝑧 goes around the contour,
divided by 2𝜋, will equal the number of roots of the numerator inside the contour, i.e. the number
of unstable roots of the characteristic polynomial of the RC system. If there is no encirclement of
the origin, the RC system is asymptotically stable.
The characteristic polynomial can be subtracted from one to produce
𝑃𝑐 (𝑧) = 1 − 𝑃(𝑧) = 𝑧 −2𝑝 𝐻(𝑧)(𝛼1 𝑧 𝑝 + 𝛼2 )(1 − 𝜙𝐹(𝑧)𝐺(𝑧)) = 0 (8)
The polar plot of 𝑃𝑐 (𝑧) is identical to 𝑃(𝑧) except that is has been shifted by one unit along the real
axis, and the sign has been switched. Then the system is stable if the polar plot of 𝑃𝑐 (𝑧) does not
encircle the point (+1,0). Typically, 𝑝 is a large number. The Nyquist contour at infinity maps onto
the origin, and the branch cut goes back and forth along the real axis from 𝑃𝑐 (−1) at Nyquist fre-
quency to the origin. A polar plot can be utilized as a graphical approach to investigate the true
stability boundary of RC systems as suggested in Reference 6, by zooming into the crucial part of
the figure and counting the number of encirclements of the point (+1,0) by visualization. This can
quickly become impractical as the number of samples per period 𝑝 increases, which can make it
difficult to count the number of encirclements. In order to resolve this difficulty, a computer algo-
rithm is developed based on the polar plot, adding the phase change from sample to sample of
closely sampled points to determine if the system is stable or not.
The mathematics for the computer algorithm to determine whether a second order RC system is
stable or not will be described in this section. A stability angle 𝜓 is defined between the positive

5
real x-axis direction and the vector originating from point (+1, 0) that intersects the vector 𝑟⃗ =
|𝑧|𝑒 𝑖𝜔𝑇𝑠 that traverses the polar plot as shown in Figure 5. An expression based on vector analysis
is obtained and relates stability angle 𝜓 to phase angle 𝜃 and the magnitude of 𝑧 as expressed in
Equations (9 to 12). As 𝑧 traverses the polar plot, the stability angle is wrapped for each revolution
in order to stay in the range [0° , 360° ] where 0° maps to the position when 𝜓 is aligned with the
positive x-axis. Counting the number of encirclements of the point (+1,0) is of critical importance
to determine the stability of the system. The principle of argument applied to 𝑃𝑐 (𝑧) says that the
associated second order RC system is asymptotically stable if there is no net encirclement of (+1,0),
i.e. there are no roots of the numerator polynomial in Equation (8) outside the unit circle, and there-
fore no roots of the second order RC characteristic equation. If the stability angle reaches a value
of 0° as 𝑧 traverses the polar plot, there exists an encirclement of the point (+1, 0) and the computer
algorithm will count all the clockwise and counterclockwise encirclements. Since the open loop
system 𝑅(𝑧)𝐺(𝑧) is designed to be stable, i.e. no poles outside the unit circle, a net clockwise
encirclement makes the system unstable. If the stability angle variations do not reach a value of
0° as 𝑧 traverses the polar plot, the point (+1, 0) is not encircled and the RC system will be consid-
ered as stable.
𝑟⃗ = |𝑟⃗|𝑒 𝑖𝜔𝑇𝑠 = |𝑟⃗|𝑒 𝑖𝜃 = 𝑆⃗ + 𝑖⃗ ⇒ 𝑆⃗ = 𝑟⃗ − 𝑖⃗ (9)

𝑆⃗ = (|𝑟⃗| cos 𝜃 − 1)𝑖⃗ + (|𝑟⃗| sin 𝜃)𝑗⃗ = 𝑆𝑥 𝑖⃗ + 𝑆𝑦 𝑗⃗ (10)

𝑆⃗ ∙ 𝑖⃗ |𝑟⃗| cos 𝜃 − 1
cos 𝛼 = = (11)
|𝑆⃗||𝑖⃗| √|𝑟⃗|2 cos 2 𝜃 − 2|𝑟⃗| cos 𝜃 + 1 + |𝑟⃗|2 sin2 𝜃

|𝑟⃗| cos 𝜃 − 1
𝜓 = cos−1 ( ) (12)
√|𝑟⃗|2 − 2|𝑟⃗| cos 𝜃 + 1
Figure 6 shows an example of a stable system where the Nyquist contour does not enclose the
point (+1, 0) and the stability angle starts out at 180° corresponding to a phase angle of 0° . The
stability angle increases until it reaches a peak at approximately 210° before it starts decreasing
until it reaches a minimum at approximately 149° . The phase angle variations are not shown but
consist of a clockwise rotation between 0° and 360° before the angle is wrapped to start another
revolution. Angular variations are identical for each of the four considered revolutions around the
polar plot since the compensator in this example was chosen to be an exact inverse of the feedback
system transfer function (this applies to any system with a stable inverse 𝐺 −1 (𝑧).
Figure 7 shows an example of an unstable system where the polar plot encloses the point (+1,
0) and the stability angle starts out at 180° corresponding to a phase angle of 180° . The stability
angle decreases until it reaches a minimum of 0° , indicating a clockwise encirclement of (+1, 0),
before it changes to 360° as it enters the fourth quadrant and continues to decrease until it reaches
the starting position at 180° . Phase angle variations are not shown but consist of a clockwise rota-
tion that is identical to the stability angle variations since origin points for both of these angles,
(0,0) and (+1,0), are enclosed inside the polar plot. Angular variations are repeated for each of the
four considered revolutions around the polar plot since the compensator in this example was also
chosen to be an exact inverse of the feedback system transfer function. To highlight the benefits of
the algorithm, a system with a compensator that is not an exact inverse and an increased period
(𝑝 = 20) was considered in Figure 8. It can be impractical to graphically determine whether the
system is stable or not, especially if there are both clockwise and counter clockwise curves.

6
There are various methods to design the compensator 𝐹(𝑧). Initially, the ideal compensator is
the system inverse, 𝐹(𝑧) = 𝐺 −1 (𝑧), but this approach can only be used if the inverse is stable. This
is rarely the case because when a continuous system is fed by a zero-order hold and the equivalent
z-transfer function 𝐺(𝑧) is generated, usually there are zeros introduced outside the unit circle. For
simplicity, a base case is used here for evaluating stability that comprises only systems that have a
stable inverse, making the product of 𝐹(𝑧)𝐺(𝑧) = 1. The period is picked as 𝑝 = 4 time steps,
sample frequency as 𝑓𝑠 = 100𝐻𝑧 and no cutoff filter is used, i.e. 𝐻(𝑧) = 1. The computer algo-
rithm described in this section was developed and run for different second order RC weights 𝛼2 ,
and the RC gain 𝜙 was varied until the stability boundary was detected, i.e. when the stability angle
𝜓 reached a minimum close to 0° . Note the computer algorithm implemented a tolerance value of
0.01° for the calculated stability angle when searching for the true stability boundary. The maxi-
mum RC gains represent the true stability boundary and are shown in Figure 9. There is a strong
dependency between the maximum RC gains and the RC weights. For positive weights (𝛼2 > 0),
the variation in stability boundary is drastic and the maximum gain is reached at a weight of
𝛼2 ~0.33. In contrast for negative weights (𝛼2 < 0), the system is more bounded since the variation
range of RC gain changes is less drastic.
One might wonder, if it is possible to mathematically determine the true stability boundary, why
it can’t be used as a way to design RC systems. One reason lies in the computational penalty asso-
ciated with this approach. For instance, each data point in Figure 9 consumed about 1 to 2 hours of
computational time using a fine sampling. Performing such computations for a control system will
be impractical and necessitates development of sufficient conditions that can simplify the RC sys-
tem design process.
SECOND STABILITY CONDITION
This section reviews a sufficient condition for asymptotic stability of second order RC systems
that was derived using a rigorous mathematical approach in Reference 6. The sufficient criterion
for stability |𝐻(𝑧)[1 − 𝜙𝐹(𝑧)𝐺(𝑧)]| < 1 was developed for first order RC. The distinction be-
tween the true stability boundary and this sufficient condition for first order RC was shown to be
of no practical significance in Reference 3. The derivation was then extended to make an analogous
higher order RC sufficient condition (for second order) requiring satisfaction of the following ine-
quality for all frequencies 𝜔 up to Nyquist
|𝐻(𝑧)[1 − 𝜙𝐹(𝑧)𝐺(𝑧)][𝛼1 𝑧 𝑝 + 𝛼2 ]| < 1 (13)
It was assumed that the characteristic polynomial for feedback system 𝐺(𝑧), the compensator 𝐹(𝑧),
and the cutoff filter 𝐻(𝑧) have all roots inside the unit circle. The derivation uses the stability
analysis related to Equations (7) and (8), and notes that if |𝑧| < 1 for all 𝜔 then the contour involved
cannot encircle the point (+1,0).
Reference 4 introduced a negative weight to benefit applications where the period has random
fluctuations. In this case this sufficient condition becomes restrictive and is far from the stability
boundary. Figure 10 shows the relationship between the maximum value of RC gain 𝜙 for stability
based on Equation (13), and it is given as a function of RC weight 𝛼2 . The left plot represents
second order RC designs that can have a compensator that is modeled as an exact inverse of the
feedback system. The design used for generating the right plot includes the third order model of a
commercial robot link described in Equation (6) and an eight gain FIR compensator design using
the method of Reference 11. If 𝛼2 is chosen positive, the maximum gain values that guarantee
asymptotic stability are constant. If 𝛼2 is negative, the stability boundary maximum gains are ap-
proximated very well but there is a gap in approximating lower gains when compared to the true
stability boundary. The condition becomes more restrictive as 𝛼2 is decreased further. The design

7
that implemented an 8 gain FIR compensator instead of an inverse, given in the right plot, becomes
even more restrictive when weights are negative.
THIRD STABILITY CONDITION
This section presents a new sufficient condition for second order RC systems by paralleling the
development of a stability condition for first order RC in Reference 2 and using a partial fraction
formulation similar to References (9) and (10). We delete the cutoff filter, making 𝐻(𝑧) = 1, in the
following analysis. Often one designs a RC system with this assumption. Then when applying the
control law in hardware, one may need to create a cutoff of the learning in order to prevent insta-
bility from high frequency model error. Since one presumably does not know this error, the cutoff
filter is tuned in hardware after observing any unstable growth of error, and determining its fre-
quency content. This allows the repetitive control transfer function in Equation (3) to be expressed
in partial fraction form as
𝑈(𝑧) 𝜙𝐹(𝑧)(𝛼1 𝑧 𝑝 + 𝛼2 ) 𝐴 𝐵
𝑅(𝑧) = = 2𝑝 𝑝
=( 𝑝 + 𝑝 ) 𝜙𝐹(𝑧)
𝐸(𝑧) 𝑧 − 𝛼1 𝑧 − 𝛼2 𝑧 − 1 𝑧 + 𝛼2
(14)
1 −𝛼22
𝐴= , 𝐵= , 𝛼1 + 𝛼2 = 1
1 + 𝛼2 1 + 𝛼2
Note that using 𝛼1 = 2 and 𝛼2 = −1 results in repeated roots for 𝑅(𝑧) on the unit circle and makes
𝐴 approach ∞, but creates a zero slope at the bottom of all notches. Reference 4 suggests weights
𝛼1 = 1.85 and 𝛼2 = −0.85 in order to widen the notches in second order RC. The block diagram
for the second order RC using the partial fraction form of 𝑅(𝑧) gives the parallel structure shown
in Figure 11. The characteristic polynomial can then be expressed using the parallel structure as
𝐴 𝐵
1 + 𝑅(𝑧)𝐺(𝑧) = 1 + ( + ) 𝜙𝐹(𝑧)𝐺(𝑧) = 0 (15)
𝑧 𝑝 − 1 𝑧 𝑝 + 𝛼2
The Nyquist stability criterion can be used to determine asymptotic stability of the system. If there
are no encirclements of the origin, the RC system is stable. Going around the unit circle, the term
(𝑧 𝑝 − 1) in the denominator presents 𝑝 poles on the unit circle and the term (𝑧 𝑝 + 𝛼2 ) has all roots
inside the unit circle if −1 < 𝛼2 < 1. Because of p poles on the unit circle, and the need to go
around each of these poles with the Nyquist contour in the left plot of Figure 4, this approach is
inconvenient. To avoid the difficulty with implementation of Nyquist stability criterion, Equation
(15) is multiplied by (𝑧 𝑝 − 1) and expressed as
𝑧𝑝 − 1
𝑧 𝑝 − [1 − (𝐴 + 𝐵 ) 𝜙𝐹(𝑧)𝐺(𝑧)] = 0 (16)
𝑧 𝑝 + 𝛼2
The numerator is the characteristic polynomial of interest for the RC system. The denominator
contains only poles inside the unit circle, by assumption that 𝐺(𝑧), 𝐹(𝑧) have no poles outside the
unit circle and considering −1 < 𝛼2 < 1, the term (𝑧 𝑝 + 𝛼2 ) has all its roots inside the unit circle.
The difficulty in application of Nyquist stability criterion, presented by going around 𝑝 poles on
the unit circle, is removed. If the phase of the left-hand side in Equation (16) comes back to the
original value after 𝑧 goes around the contour, then there are no roots of the characteristic polyno-
mial inside the contour in the right plot in Figure 4 and the system is asymptotically stable.
The same statement still applies if Equation (16) is multiplied by 𝑧 −𝑝 , which introduces 𝑝 poles
at the origin. These poles are outside the Nyquist contour and do not influence stability. The refor-
mulated characteristic polynomial is

8
𝑧 𝑝 −1
1 − 𝑧 −𝑝 [1 − (𝐴 + 𝐵 𝑧𝑝 +𝛼 ) 𝜙𝐹(𝑧)𝐺(𝑧)] = 0 (17)
2

There are more poles than zeros in Equation (17), the part of the contour can be ignored out at
infinity since it maps onto the origin, and the branch cut goes back and forth along the real axis
from the origin to the value at the Nyquist frequency. The repetitive control system is asymptoti-
cally stable if and only if the mapping of 𝑧 going around the unit circle does not go around the
origin. Because |𝑧 −𝑝 | = 1 for 𝑧 on the unit circle
1 𝛼2 𝑧 𝑝 −1
|1 − (1+𝛼 − 1+𝛼2 𝑝 ) 𝜙𝐹(𝑧)𝐺(𝑧)| <1 (18)
2 2 𝑧 +𝛼2

𝑧 𝑝 −1
is a sufficient condition for stability. Since 1 − (𝐴 + 𝐵 ) 𝜙𝐹(𝑧)𝐺(𝑧) could not go around the
𝑧 𝑝 +𝛼2
origin, there is no need to consider encirclements of -1 (or +1) for this criterion.
In order to evaluate the sufficient stability condition in Equation (18), the same base case from
the previous section that comprises only systems that have a stable inverse is used here. The RC
gain 𝜙 is varied to test stability. Figures 12 and 13 plot the “Stability Magnitude”, i.e. the left-hand
side of Equation (18) for different second order RC weights between −1 < 𝛼2 < 1 and for a fre-
quency range that varied from 0 Hz to Nyquist at 50 Hz. For negative weights in Figure 12, the
stability magnitude starts at a value just below 1 when the RC gain is 0. As the RC gain is increased,
the stability magnitude decreases initially with downward pointing peaks appearing at fundamental
and harmonics. As the RC gain is increased further, the stability magnitude starts to increase with
peaks pointing upward instead. The system becomes unstable when the stability magnitude values
at the peaks exceed unity. A similar progression occurs for positive weights in the range 0 < 𝛼2 <
0.5 as shown in Figure 13 but at frequencies halfway between fundamental and harmonics or be-
tween adjacent harmonics. As the range of the second order RC weights increases to 0.5 < 𝛼2 <
1, the stability magnitude immediately exceeds 1, the sufficient condition is not satisfied. This
result prevents use of this particular sufficient condition for 𝛼2 in this range.
Inequality in Equation (18) represents a necessary and a sufficient condition for asymptotic sta-
bility of second order RC system for all possible values of period 𝑝 and should be satisfied for all
frequencies from zero to Nyquist. Results of an evaluation of the magnitude of the stability condi-
tion for three second order RC weights and two values of period (𝑝 = 4 and 𝑝 = 50) are shown
in Figure 14. Besides the difference in the number of peaks resulting from a higher frequency sig-
nal, the magnitude of the stability condition is identical at the peaks for these considered scenarios.
This indicates that Equation (18) can be used to assess the stability of a system over the range of
frequencies from 0 to Nyquist independent of the number of samples in a period 𝑝.
The relationship between RC gain 𝜙 and RC weights 𝛼2 is shown in Figure 15. The value of
gain 𝜙 that guarantees system stability is linearly related to RC weights when the range is chosen
as −1 < 𝛼2 < 0.5. As the RC weight 𝛼2 exceeds 0.5, the sufficient condition does not converge to
a positive RC gain that can predict the system to be stable and thus RC gains were allocated a value
of 0. This sudden change in prediction is instigated by the change in the magnitude of the term
inside parentheses in Equation (18), more specifically the term 𝛼22 (𝑧 𝑝 − 1) becomes greater than
𝑧 𝑝 + 𝛼2 , causing the stability condition magnitude to exceed unity, and resulting in no conclusion
about gains that could produce stability.
FOURTH STABILITY CONDITION
This section develops a sufficient condition based on the third stability condition in the previous
section by eliminating dependence on period 𝑝 and thus simplifying the condition for ease of

9
evaluation. For designing a second order RC system, one is generally interested in finding a condi-
tion that is not too conservative and guarantees stability for all possible weights and values of period
𝑝. This condition should also be satisfied for all frequencies from zero to Nyquist frequency. For
this purpose, the sufficient stability condition in Equation (18), that relied on closed loop system
representations in Figures 1 and 2, will be simplified to result in a fourth stability condition that
can be used for second order RC designs.
Consider the closed-loop system shown in Figure 1 and combine it with the second order RC
design block shown in Figure 2. The closed loop system is asymptotically stable if Equation (18)
is satisfied for all 𝑧 = 𝑒 𝑖𝜔𝑇𝑠 and for all frequencies to Nyquist, 0 ≤ 𝜔 ≤ 𝜔𝑁 . In order to simplify
the equation development below, a transfer function 𝑊(𝑧) is defined as
1 𝑧 𝑝 −1
𝑊(𝑧) = (1 − 𝛼22 ) (19)
1+𝛼2 𝑧 𝑝 +𝛼2

The frequency response of the stability condition is developed as follows, 𝑀(𝜔) =


𝑀𝐹 (𝜔)𝑀𝐺 (𝜔), 𝑒 𝑖𝜃(𝜔) = 𝑒 𝑖(𝜃𝐹 (𝜔)+𝜃𝐺 (𝜔)) , |1 − 𝜙 𝑊(𝜔)𝑀(𝜔)(cos 𝜃(𝜔) + 𝑖 sin 𝜃(𝜔))| < 1.
The magnitude of frequency response is calculated followed by squaring both sides to result in the
following expressions
√1 − 2 𝜙 𝑊 𝑀 cos 𝜃 + 𝜙 2 𝑊 2 𝑀2 cos 2 𝜃 + 𝜙 2 𝑊 2 𝑀2 sin2 𝜃 < 1
−2 𝜙 𝑊 𝑀 cos 𝜃 + 𝜙 2 𝑊 2 𝑀2 < 0
(20)
cos(𝜃𝐹 (𝜔) + 𝜃𝐺 (𝜔))
0 < 𝜙 < 2 ( min )
0≤𝜔≤𝜔𝑁 𝑊(𝜔) 𝑀(𝜔)
Since 𝜙 > 0, angle 𝜃(𝜔) ∈ [−90° , 90° ], i.e. 𝜃(𝜔) lies in either first or fourth quadrant, Equation
(20) can be simplified further by finding the max(𝑊(𝜔)). Equation (21) shows the magnitude for
𝑊(𝑧) over the entire frequency range.
1 𝑒 𝑖𝜔𝑝𝑇𝑠 −1 |𝑒 𝑖𝜔𝑝𝑇𝑠 −𝛼2 𝑒 𝑖𝜔𝑝𝑇𝑠 +𝛼2 |
|𝑊(𝑧)|𝑧=𝑒 𝑖𝜔𝑇𝑠 = | (1 − 𝛼22 𝑒 𝑖𝜔𝑝𝑇𝑠 +𝛼 )| =
1+𝛼 2 2 |𝑒 𝑖𝜔𝑝𝑇𝑠 +𝛼2 |
(21)
1−2𝛼2 +2𝛼22 +2𝛼2 cos(𝜔𝑝𝑇𝑠 )(1−𝛼2 )
= √ 1+𝛼22 +2𝛼2 cos(𝜔𝑝𝑇𝑠 )

The magnitude function 𝑊(𝜔) attains a maximum at 𝜔𝑇𝑠 = 0° for all RC weights in the range
−1 < 𝛼2 < 0.5. Equation (22) shows the maximum magnitude which occurs at 𝜔𝑇𝑠 = 0° .

1 1
max (𝑊(𝜔)) = √ 2 = (22)
0≤𝜔≤𝜔𝑁 1 + 𝛼2 + 2𝛼2 1 + 𝛼2

By combining the result in Equations (20) and (22), a new simplified sufficient stability condition
is obtained
cos 𝜃(𝜔)
0 < 𝜙 < 2 ( min ( )) (1 + 𝛼2 )
0≤𝜔≤𝜔𝑁 𝑀(𝜔) (23)
° °
𝜃(𝜔) = 𝜃𝐶 (𝜔) + 𝜃𝐺 (𝜔), 𝜃(𝜔) ∈ [−90 , 90 ], −1 < 𝛼2 < 0.5
During development of this simplified condition, no additional assumptions were made to Equation
(18), and thus both the third and the fourth sufficient conditions provide identical results. The linear
relationship between RC gain and weight shown in Figure 15 was obtained using Equation (23).

10
The simplified fourth condition provides advantages in ease of evaluation and eliminates the de-
pendence on period 𝑝 in the stability expression.
COMPARISON BETWEEN TRUE BOUNDARY AND SUFFICIENT CONDITIONS
To illustrate the practical applications in RC design, a comparison between the true stability
boundary and the sufficient conditions is necessary. The first, second and fourth stability conditions
are compared in Figures 16a and 16b. An evaluation for a base case that comprises systems with
exact stable inverses is provided in Figure 16a while a third order transfer function model of a robot
link described in previous sections and compensated with an 8 gain FIR design is shown in Figure
16b. In general, it is found that neither condition is superior when compared to the first stability
condition, i.e. the true stability boundary, over the entire range of weights. RC designers are left
with the choice to breakup stability regions into different zones based on the proximity between
sufficient condition predictions and the true stability boundary. Consequently, the second stability
condition (solid) is the only sufficient condition that can predict non-zero gains producing stability
in the range of weights between 0.5 and 1, and it coincides with the true stability boundary (round)
when 𝛼2 = 1. For weights between 0 and 0.5, the third or fourth stability conditions (dashed) best
approximate the true stability boundary. From -1 to 0, the second stability condition (solid) approx-
imates the maximum gains very well but falls short of predicting lower gains of zero. Instead, the
third or fourth stability conditions can be used between -0.7 and 0. All sufficient conditions have
either a gap or become too conservative between -1 and -0.7. Thus, one needs to use multiple suf-
ficient conditions based on the weights considered for each design.
To realize the practical implementation of the different stability conditions, two second order
RC designs are given in Figure 17 that correspond to weights of 𝛼1 = 1.85 and 𝛼2 = −0.85 in left
plot and weights of 𝛼1 = 0.5 and 𝛼2 = 0.5 in right plot. For simplicity, the period addressed is set
to 10-time steps. The compensator used here is an exact system inverse. The values of gain 𝜙 were
obtained from Figure 16a as the maximum allowed for stability according to each condition or
boundary. The solid results reflect gains obtained from true stability boundary, dashed-dotted re-
sults correspond to the second stability condition, and dashed results use gains from the third or
fourth stability condition. In general, the result obtained with gains from third or fourth stability
condition captures the magnitude of the sensitivity transfer function at the addressed frequencies
but misses on amplification characteristics between addressed frequencies. The discrepancy in am-
plification amplitude in comparison to result using gain from true stability boundary is greater when
using negative weights (left plot on Figure 17). Note that one needs to be careful when using this
condition with negative weights, the influence of disturbances between addressed frequencies by
the RC system can be easily overlooked, since the true stability boundary is not typically known.
Similarly, the magnitude obtained with gains from the second stability condition capture magni-
tudes at addressed frequencies fairly well with differences in the amplification amplitude at non-
addressed frequencies. It is worth noting, the upper bound of maximum gains was used for the
second stability condition and hence the results are close to the true stability boundary when using
negative weights. For positive weights, the magnitude using the second condition is identical for
all weights since the maximum gain is constant.
To gain insight into the variation of weights, the magnitudes of the sensitivity transfer function
for three second order RC designs are given in Figure 18. These designs correspond to weights of
𝛼1 = 1.85 and 𝛼2 = −0.85, 𝛼1 = 1.25 and 𝛼2 = −0.25, 𝛼1 = 0.5 and 𝛼2 = 0.5. The left plots
use a gain of 𝜙 = 0.5 and right plots use a gain of 𝜙 = 1. The compensator was assumed to be an
exact system inverse. The Bode integral theorem or waterbed effect is evident in both plots. If the
frequency response magnitude is near zero in a certain frequency range (addressed frequencies),
substantially attenuating the influence of the disturbances in this range, then there must be some

11
other frequency range (unaddressed frequencies) for which the magnitude is greater than one and
disturbances are amplified. Note that decreasing weight 𝛼1 (increasing 𝛼2 ) reduces the amplifica-
tion at unaddressed frequencies and narrows the valleys at addressed frequencies. Hence, when
selecting weights for second order RC designs with fluctuating periods, a designer is faced with a
tradeoff decision between width of valleys at addressed frequencies and the amplification of dis-
turbances at unaddressed frequencies.
Another important aspect of designing RC systems is the convergence rate. To investigate the
performance of second order RC from the point of view of convergence rate as a function of gains
and weights, two designs that correspond to 𝛼1 = 1.85 and 𝛼2 = −0.85, 𝛼1 = 0.5 and 𝛼2 = 0.5
are considered and the gains are set to 𝜙 = 0.5 and 𝜙 = 1. The same third order model as in pre-
vious sections is used with an 8 gain FIR compensator. A periodic disturbance of period one second
is added to the output of the feedback system. The disturbance is formed by including sinusoids at
every integer frequency from 0 to 49 Hz, each with an amplitude of 0.05, and with a phase angle
that is chosen randomly as in Reference 6. The task of the RC system is to eliminate this output
disturbance. Figure 19 gives the resulting RMS error for each period or repetition, for each design.
An assessment of curve pairs with the same weights, shows that the higher gains have a faster
learning rate. Since the boundary of a sufficient stability condition is not necessarily the actual
stability boundary, the RC may still learn well at this boundary. Comparison of curves with the
same gain indicates that decreasing weight 𝛼1 (increasing weight 𝛼2 ) makes the system learn faster.
There is a tradeoff, decreasing the gain 𝜙 limits the amplification between addressed time steps,
but creates narrower notched that result in more sensitivity to fluctuations of the period.

CONCLUSION
Second order RC systems can be designed to reduce sensitivity to period fluctuations or impre-
cise knowledge of the period. Evaluating stability of such systems becomes complicated, especially
for systems designed with negative weights, making it hard to design second order RC systems.
This paper addressed this challenge by evaluating four different approaches:
− A computer algorithm that applies to higher order RC systems with positive or negative
weights, and samples very frequently adding up the phase changes from sample to sample. This
approach determines the first stability condition, i.e. the true stability boundary, that guarantees
asymptotic stability. One disadvantage is the heavy computational penalty.
− A rigorous mathematical approach from a previous publication, that is analogous to the funda-
mental necessary and sufficient condition for first order RC, provides a sufficient condition for
second order RC, the second stability condition. A disadvantage for this approach, it misses a
region of lower gains for stability when the weights are negative.
− A partial fraction formulation that utilizes a reformulated Nyquist criterion and parallels the
development of a sufficient condition for first order RC, provides sufficient third stability con-
dition for second order RC. One disadvantage for this approach is lack of identifying any pos-
sible stability for RC weight greater than 0.5, limiting its application to other ranges.
− A simplification of the third stability condition eliminates dependence on period resulting in
the fourth sufficient condition that is much easier to apply. Evaluation of the third and the fourth
stability conditions provides identical results suggesting interchangeability. This condition fills
in much of the range of lower gains missing in the second stability condition, but the maximum
value is best only from 0 to 0.5.
A comparison of when to use each stability condition based on RC design weight α2 was per-
formed. From 0.5 to 1, the second stability condition (solid) is the only sufficient condition that
allows a non-zero gain, and it is not too far from the true stability boundary. From 0 to 0.5, the third

12
and fourth stability conditions (dashed) best approximate the true stability boundary. For designs
in the negative weight range from -1 to 0, which are required for widening notches, the maximum
gains available are best approximated using the second condition (solid), closely matching the if-
and-only-if condition. However, the second condition leaves a hole in the stability region for low
gains. The third or fourth stability conditions (dashed) fill this hole between -0.7 and 0. From -1 to
-0.7, the sufficient conditions collectively leave a gap on the stability region for intermediate gains.
In summary, this investigation suggests one needs to use multiple sufficient conditions that can be
good approximations to true stability boundary based on the range of weights being considered.

REFERENCES
1. E. S. Ahn, R. W. Longman, J. J. Kim, and B. N. Agrawal, “Evaluation of Five Control Algo-
rithms for Addressing CMG Induced Jitter on a Spacecraft Testbed,” The Journal of the Astro-
nautical Sciences, Vol. 60, Issue 3, 2015, pp. 434-467.
2. 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.
3. T. 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.
4. M. Steinbuch, “Repetitive Control for Systems with Uncertain Period-Time,” Automatica, Vol.
38, No. 12, 2002, pp. 2103-2109.
5. 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.
6. C.-P. Lo and R. W. Longman, “Frequency Response Analysis of Higher Order Repetitive Con-
trol,” Advances in the Astronautical Sciences, Vol. 123, 2006, pp. 1183-1202.
7. H.-J. Guo, R. W. Longman, T. Ishihara, “A Design Approach for Insensitivity to Disturbance
Period Fluctuations Using Higher Order Repetitive Control,” Proceedings of the 19th World
Congress of the International Federation of Automatic Control, Cape Town, South Africa, Au-
gust 24-16, 2014.
8. E. S. Ahn, R. W. Longman, and J. J. Kim, “Comparison of Multiple-Period and Higher Order
Repetitive Control Used to Produce Robustness to Period Fluctuations,” Advances in the As-
tronautical Sciences, Vol. 148, 2013, pp. 179-202.
9. Peiling Cui, Zhiyuan Liu, Guoxi Zhang, Han Xu, and Richard W. Longman, “A Novel Second
Order Repetitive Control that Facilitates Stability Analysis and its Application to Magnetically
Suspended Rotors,” IEEE Access, Vol. 7, 2019, pp. 149857-149866.
10. Peiling Cui, Ayman Ismail, Richard W. Longman, Zhiyuan Liu, and Han Xu, “Theory and
Evaluation of a Stability Condition for Second Order Repetitive Control,” submitted to
AIAA/AAS Spaceflight Mechanics Conference, Orlando FL, January 2020.
11. 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.

Figure 1. Basic Block Diagram of a General RC System

13
Figure 2. Block Diagram of a Standard Second Order RC

Figure 3. Magnitude of Sensitivity Transfer Function for 12 Gain FIR Design


(First Order RC: Solid, Second Order RC: Dashed, RC Gain = 0.1: Thick, RC Gain = 1: Thin)

Im(z) Im(z)

Re(z) Re(z)

Figure 4. Nyquist Contour: Going Around 𝒑 Poles on Unit Circle (Left), Going Around Unit Cir-
cle Without 𝒑 Poles on Unit Circle (Right)

14
Figure 5. Vector Relations for Defining Stability Angle

Figure 6. Stable Second Order RC System with Exact System Inverse Compensator (𝒑 = 𝟒)
(Polar Plot: Left, Stability Angle: Right)

Figure 7. Unstable Second Order RC System with Exact System Inverse Compensator (𝒑 = 𝟒)
(Polar Plot: Left, Stability Angle: Right)

15
Figure 8. Second Order RC of a 3rd Order System with 8 Gain FIR Compensator (𝒑 = 𝟐𝟎)
(Polar Plot: Left, Stability Angle: Right)

Figure 9. Relationship Between 𝝓 and 𝜶𝟐 Based on True Stability Condition


(Exact System Inverse Compensator: Left, 8 Gain FIR Compensator Design: Right)

Figure 10. Relationship Between 𝝓 and 𝜶𝟐 Based on Second Stability Condition

Figure 11. Block Diagram for a Parallel Structure Second Order RC

16
Stability Magnitude

Frequency (Hz)
Figure 12. Magnitude of Third Stability Condition for Different Weights (−𝟏 < 𝜶𝟐 < 𝟎)
Stability Magnitude

Frequency (Hz)
Figure 13. Magnitude of Third Stability Condition for Different Weights (𝟎 < 𝜶𝟐 < 𝟏)

17
Stability Magnitude

Frequency (Hz)
Figure 14. Magnitude of Third Stability Condition for Different Periods (𝒑 = 𝟒, 𝒑 = 𝟓𝟎)

Figure 15. Relationship Between 𝝓 and 𝜶𝟐 Based on Third or Fourth Stability Condition
(Exact System Inverse Compensator: Left, 8 Gain FIR Compensator Design: Right)

18
Figure 16a. Comparison of Second Order RC Stability with Exact Inverse Compensator
(True Boundary: Circles, Second Stability Condition: Solid, Fourth Stability Condition: Dashed)

Figure 16b. Comparison of Second Order RC Stability with 8 Gain FIR Compensator
(True Boundary: Circles, Second Stability Condition: Solid, Fourth Stability Condition: Dashed)

19
Figure 17. Sensitivity Transfer Function of Second Order RC Systems with Different Gains
(Left: 𝜶𝟏 = 𝟏. 𝟖𝟓, 𝝓 = 𝟏. 𝟖𝟓: Solid, 𝝓 = 𝟎. 𝟐𝟗: Dashed, 𝝓 = 𝟏. 𝟑𝟎: Dashed-Dot)
(Right: 𝜶𝟏 = 𝟎. 𝟓, 𝝓 = 𝟐. 𝟗𝟓: Solid, 𝝓 = 𝟐. 𝟖𝟓: Dashed, 𝝓 = 𝟐: Dashed-Dot)

Figure 18. Sensitivity Transfer Function of Second Order RC Systems with Different Weights
(𝝓 = 𝟎. 𝟓: Left, 𝝓 = 𝟏: Right, 𝜶𝟏 = 𝟐: Dashed, 𝜶𝟏 = 𝟏. 𝟐𝟓: Dashed-Dot, 𝜶𝟏 = 𝟎. 𝟓: Solid)

Figure 19. Convergence Rates for 2nd Order RC, 3rd Order System with 8 Gain FIR Compensator
(𝜶𝟏 = 𝟏. 𝟖𝟓, 𝝓 = 𝟏: Solid, 𝜶𝟏 = 𝟏. 𝟖𝟓, 𝝓 = 𝟎. 𝟓: Dashed, 𝜶𝟏 = 𝟎. 𝟓, 𝝓 = 𝟏: Dashed-Dotted,
𝜶𝟏 = 𝟎. 𝟓, 𝝓 = 𝟎. 𝟓: Dotted)

20

You might also like