Vous êtes sur la page 1sur 5

AA448 - CONTROL IN AEROSPACE SYSTEMS LAB, WINTER 2013 1

Discussion of Topics Covered in Control in Aerospace Systems


Taylor Campbell, Jaime Katzer, and Kelsey Moser
AbstractThis course provided an introduction to the
eld of controls in aerospace. The rst focus was on
creating a mathematical model of a real system, of varying
levels of accuracy; the second, on solving the model to
predict future behavior of the system, and the third, on
measuring the behavior of the system and adding adding
a controller to modify the output to match either an input
or a desired set of metrics.
I. INTRODUCTION
C
ONTROLS in aerospace has a simple goal: model a
system, solve the model to predict the future behavior
of the system, and change the behavior of the model to match
design and performance specications without modifying the
plant itself. Achieving these goals, however, is not simple. A
portion of the Control in Aerospace Systems course, AA447,
was devoted to an introduction in each of these topics, with
an end goal of making a controlled plant meet performance
criteria the original system cannot. The arc of the class is
expressed in the gure below, and will be dissected and
explained throughout this paper.
Fig. 1. Governing diagram for AA447 course
II. MODELING
The diagram begins with a real system, affected by intrinsic
properties and by any number of inuences from the outside
world. In order to successfully modify the way the system
behaves, it must rst be modeled so that its behavior can be
simulated and future behavior can be predicted. System model-
ing necessitates approximation, because most real systems are
T. Campbell, J. Katzer, and K. Moser are undergraduate students in
the William E. Boeing Department of Aeronautics and Astronautics at the
University of Washington, Seattle, WA, USA
too complicated and affected by too many nuances to represent
with complete accuracy, or if the entire system and inputs can
be perfectly modeled, the math is likely to be difcult and
time consuming to process. There is a trade off between how
accurately you model the system and the ease with which you
do it. [1]
A. Modeling Mechanical Systems
Fig. 2. Traditional spring-mass-damper system
Simple mechanical systems provide a logical jumping off
point to system modeling. Applying Newtons second law,
F = ma, allows a simple system (like a spring-mass-damper
found above in Fig. 2) to be represented as a linear ordinary
differential equation (ODE) called the equation of motion
(EOM). The EOM representing the system in Fig 2. is found
below.
m x(t) + c x(t) + kx(t) = u(t)
The mass of the system in this case is m, k is the linear
spring constant, and c is the linear damping constant. x(t)
is the position of the mass, and u(t) is the external force
acting on the system. (Mechanical systems in rotation instead
of translation heed similar rules, with moment of inertia,
J, used instead of mass, and angular position, (t) used
instead of translational position.) Because this is a second
order linear ODE, a relatively simple analytical solution exists.
Alternatively, nonlinear models of both springs and dampers
could also be used, and may result in a more accurate system
model, but one that is more difcult to solve. Fidelity of the
model often must be sacriced to allow for a more manageable
solution a benet of the spring-mass-damper system is that
little accuracy is lost by the approximations made in the
derivation of the linear ODE model.
B. Modeling Electrical Systems
Simple electrical systems can be modeled with second
order linear ODEs in a very similar way to spring-mass-
damper systems. An equation of motion can be derived for
the current in an electrical circuit using Kirchoffs Voltage
and Current Laws, similar to the use of Newtons 2nd Law
AA448 - CONTROL IN AEROSPACE SYSTEMS LAB, WINTER 2013 2
in mechanical systems. An inductor in an electrical system
behaves similarly to a mass in a mechanical system. Likewise,
a capacitor stores and releases energy in much the same way
a mechanical spring does. A resistor in the circuit dissipates
energy, like a mechanical damper. Its important to note,
however, that while the differential equation is written in terms
of current in the system, the ODE of an electrical system
often includes both derivatives and integrals of the i(t) term,
making current the mathematical equivalent of velocity in a
mechanical system.
III. SOLVING AND SIMULATING MODELS
Once a system is modeled with a differential equation, it
must be solved to nd a single equation for the output with
respect to time. This course explored two primary ways to
solve a systems equation of motion: analytical solutions and
numerical solutions.
A. Analytical Solutions
An ODE can be solved analytically in a multitude of ways;
the primary method utilized in this course was the Laplace
transform, because it related easily to the concept of a transfer
function, to be discussed later. A brief review was also given
on the traditional methods of ODE solving.
1) Traditional Methods
By assuming an exponential solution form for a second
order linear ODE without a damping term, the following
solution for x(t) can be found, where A and are constants
depending on the initial conditions, and there is no external
forcing being applied to the system (u(t) = 0).
x(t) = Asin t +
Deriving an analytical solution in this way is difcult -
for systems larger than second order, the tabulated solutions
become unwieldy, and adding external forcing adds another
layer of complexity. Transforming an equation of motion into
the Laplace domain is another way to nd an analytical ODE
solution.
2) Laplace Transforms
The Laplace domain, denoted by an s, is sometimes called
the frequency domain. Transforming a time-domain function
to its Laplace domain equivalent can turn a complicated
differential equation solution into an exercise in algebra in
terms of multiples of X(s) and U(s). The denition of the
Laplace transform of f(t), F(s), is found below.
F(s) = L{f(t)} =

0
f(t)e
st
dt
The Laplace transform of a function exists if f(t) is piecewise
continuous and of exponential order. [1] One of the many
convenient things about converting and beginning the solution
process in the Laplace domain is that taking a derivative of a
function in the time domain is equivalent to multiplying it by
s in the Laplace domain. Transforming the equation of motion
into the Laplace domain can be done using the equation above,
or using one of the countless tables with common transforms
already calculated. After solving the resulting equation for
F(s), transforming the equation back into the time domain
will give a solution for f(t). One of the most important things
about solving ODEs with Laplace transforms is that it leads
easily to the derivation of transfer functions.
B. Numerical Solutions
In real systems, behavior does not always follow an ideal,
linear ODE. In these cases, certain simplifying assumptions
and approximations can be made, like using a spring force
proportional to the stretched length of the spring instead
of one proportional to the length cubed, or applying the
small angle approximation to discard trigonometry terms in an
equation of motion. The delity of a system model declines
as these assumptions are made; one way to avoid this is to
use numerical simulation tools, such as Simulink. Systems too
complicated to solve by analytical methods can be modeled in
Simulinks graphic interface relatively easily. As in any other
solution method, there is often a trade off between the accuracy
of the solution and the computing power used; a model using a
smaller time step yields more accurate results, but takes longer
to process.
IV. SYSTEM PROPERTIES
One of the most versatile and useful ways to dene a system
model is through its transfer function.
A. Transfer Functions
A transfer function is dened as the Laplace transform of
the output of a system, Y (s), over the Laplace transform of
the system input, U(s), with no initial conditions considered.
G(s) =
Y (s)
U(s)
=
L[output]
L[input]

ICs=0
It is a property of the system, and does not depend on the
input - instead it allows the output resulting from a given input
to be predicted. As long as the output behavior is the same,
two systems with different internal structures can have the
same transfer function; a transfer function does not describe
internal components. [1] The transfer function provides a more
succinct way to model a system in Simulink. Every transfer
function used in this course was single input, single output
(SISO). Although multi-input, multi-output (MIMO) systems
are common in control systems, they were beyond the scope
of this course. The output of a system modeled by a transfer
function in response to any input U(s) can be found by simply
multiplying the transfer function by the input in question,
making it a versatile tool for predicting system behavior across
a range of inputs.
The transfer functions dealt with in this class had polyno-
mials as both the numerator and denominator, and the roots of
these polynomials comprise the poles and zeros of a transfer
function. Zeros of a transfer function are values of s that make
the numerator equal to zero, and poles are values that cause the
denominator to be zero. A general form can be found below,
where z
x
signies a zero and p
x
denotes a pole.
G(s) =
(s + z
1
)(s + z
2
)...(s + z
m
)
(s + p
1
)(s + p
2
)...(s + p
n
)
AA448 - CONTROL IN AEROSPACE SYSTEMS LAB, WINTER 2013 3
It should be noted here that every transfer function dealt
with in this course was proper or strictly proper, meaning
the denominator was of the same or higher order than the
numerator.
1) The Importance of Poles
The poles of a function determine the stability and the
behavior of that function. In order for a system to be stable,
dened as having an output that approaches a nite number
or function, the poles must all lie in the left hand side of
the imaginary axis when plotted on the Real/Imaginary plane.
This means the real component of any pole must be negative.
It is also possible to have a stable response to an input
without a real component, where poles lie on imaginary axis.
The system, however, will not have Bounded Input Bounded
Output (BIBO) stability, dened as producing a stable output
for every input that is bounded [2] (ie. not growing without
bound, like a ramp or exponential function).
B. Performance Metrics
Given a stable system model, certain system properties can
be predicted which measure the performance of a system and
whether it meets requirements. Many of these properties deal
with how a system approaches a steady state response, or the
behavior of the system as time grows large, formally dened
as the limit as time goes to innity of the system response.
The transient response is the process of a system going from
an initial state to this steady state response. The error in steady
state, e
ss
, is dened as the difference between the system
input and the steady state response of the system, y
ss
, and is
commonly used in control systems to determine the guidance
the controller gives the plant.
The DC gain of a system is the difference in magnitude
between the steady state output of a system and the input. One
benet to modeling a system as a transfer function is that the
Final Value Theorem states that the limit of a function of t as
t approaches is equal to the limit of the Laplace transform
of the same function as s approaches zero. Finding the DC
gain of a transfer function, then, is as simple as nding this
limit. This approach can also be reversed as a simple way to
nd y
ss
of a system in response to a step input A; the steady
state response is simply the product of the step amplitude and
the DC gain.
The rise time, T
r
, of a system is the time elapsed before
the system reaches a certain percentage, , of the steady state
response, regardless of whether the system overshoots or stays
within that range. Settling time, T
s
, is similar, but it requires
the system to reach and stay within of y
ss
. Percent overshoot,
PO, is difference, in percent, between the peak response of the
system and the steady state response. The natural frequency
of a system,
n
, is the frequency that the system would
oscillate at without any external forcing, while the damped
natural frequency,
d
, is the frequency that the forced system
oscillates at.
V. DESIGNING CONTROL SYSTEMS
An open loop control system executes preprogrammed com-
mands on a schedule, without any feedback from the system
itself. A closed loop control system executes commands based
on comparisons between the system output and a desired value
of some system parameter. [1] The poles of a system dictate
the system response - this means that changing the poles of a
system changes its behavior. Only a controller with a feedback
loop, seen in Fig. 3 changes the poles of the original system
G(s), and so a feedback loop can adjust the behavior of a
system to match the desired output, unlike a simple gain or a
feed-forward loop.
Fig. 3. Feedback loop with simple gain controller [1]
The closed loop transfer function T(s) for a system with
proportional gain controller with negative unity feedback such
as that above depends on the gain K.
T(s) =
K G(s)
1 + K G(s)
This means that changing K can change the poles and thus
the behavior and stability of the system.
A. Root Locus
The easiest way to keep track of how the stability of the
system changes as the gain, K is changed is with a plot called
a root locus. As K changes, the poles of a system move - an
unstable system can become stable if its poles move from the
right to the left side of the real/imaginary plane, and in the
same token, a stable system can become unstable if pushed
with too large of a gain. A root locus maps how the poles
change and can help determine the appropriate gain to apply
via controller to a system.
To begin plotting a root locus for a system, plot the zeros
and poles of the open loop system, G(s) on the Re/Img plane.
No matter the size of your gain, the number of poles a system
has does not change. Since the root locus tracks the movement
of the poles, it makes sense that it starts at the location of
the open loop poles. While slightly less intuitive, it is also
true that the root locus ends either at open loop zeros or at
zeros at innity. The number of zeros at innity that a system
has can be determined easily, by subtracting the number of
zeros from the number of poles. They are located on root
locus asymptotes; the angle each asymptote makes, measured
counter-clockwise from the real axis is called
A
and is found
with the following equation.

A
=
(1 2k)
n m
Here, n m represents the number of poles at innity, and
k is a succession of integers starting at 0 and ending at
n m 1. These asymptotes are centered at a point
A
,
found by subtracting the sum of the locations of the system
zeros from the sum of the locations of the poles, and dividing
what remains by the number of zeros at innity. Each pole or
AA448 - CONTROL IN AEROSPACE SYSTEMS LAB, WINTER 2013 4
zero that lies on the real axis (meaning it has no imaginary
component) is numbered from right to left. Valid regions for
the root locus to lie are found to the left of odd numbered
poles or zeros. If a pole lies off of the real axis, the angle at
which it leaves to begin to chase an asymptote can be found
using the solution to the characteristic equation ((s)) of the
closed loop transfer function, T(s). The solution is a complex
number, given in terms of the poles and zeros of the system,
which means that the angle it forms on the real/imaginary
plane can be found, and that the angle of the entire solution
is equal to 2k, where k is an integer. By choosing
an s value close in magnitude to a pole off the real axis, the
angle of departure from that pole can be found by individually
adding up the angles made by the zeros and subtracting the
angles made by the poles, leaving the angle between s and the
pole in question as the unknown. [1]
B. Poles and Performance Metrics
Geometric representations of performance metrics on the
real/imaginary plane can be used to size controller gains to
meet these system requirements, if the characteristic equation
of the system is a second order polynomial or can be reason-
ably approximated as one. The angle is dened as the angle
from the negative real axis, measured in both directions, and it
is determined by the damping ratio, which is in turn a function
of the desired percent overshoot of the system. A plot showing
and the other performance metrics can be found in Fig. 4
- the necessary to satisfy PO requirements is between the
negative real axis and the wedge.
= tan
1
(

1
2

) =
ln(
PO
100
)

2
+ ln(
PO
100
)
2
Settling time (T
s
, to within a fraction ) is a function of
- poles must lie to the left of any value calculated by this
method.
=
ln()
T
s
The natural frequency,
n
dictates a semi circle centered at
the origin and of radius r that the poles must lie on, where
r =

2
+
2
d
. To restrict the damped natural frequency of
the system,
d
, poles must fall within a column extending from
the real axis to a distance
d
up and down the imaginary axis.
Fig. 4. Performance metric-restricted Re/Img plane [1]
C. System Stability
Routh Hurwitz stability criteria provide a simple and rel-
atively quick way to check a complicated system is stable.
Given a characteristic equation of a transfer function of the
form (s) = a
n
s
n
+a
n1
s
n1
+... +a
1
s +a
0
, the rst step
is to build an empty matrix with n + 1 rows, where n is the
order of the polynomial, and with h columns, where h is equal
to ceil(
n+1
2
). The rst two rows are lled in as below, starting
with the coefcient of the highest power in the upper left hand
corner, and lling in the rst two rows in a zig-zag, putting a
0 in the rightmost spot of the second row if the polynomial is
of an even power. The y row dened as the row above the k
TABLE I
ROUTH HURWITZ STABILITY MATRIX
1 2 ... h
1 an a
n2
... a
0
2 a
n1
a
n3
... 0
... ... ... z
k,i
0
n+1 z
n,1
z
n,2
z
n,i
0
row, and the x row is that above the y row. The term z
k,i
can
be found with the following equation, where k designates any
row greater than 2, and i designates the column the term is
found in. The far right column is the exception - these values
are always zero. [1]
z
k,i
=
y
1
x
i+1
y
i+1
x
1
y
1
After completing the table, the number of sign changes found
in the rst column, proceeding downwards, is the number of
poles in the right half of the real/imaginary plane. In other
words, a system is only stable if there are no sign changes in
rst column of its Routh Hurwitz table.
D. Frequency Domain Analysis
When a sine wave is input into a transfer function, the steady
state output is also a sine wave, and at the same frequency.
The output is phase shifted by , often called the phase lag,
and the amplitude is either attenuated or amplied, depending
on the system characteristics and the input.
y
ss
(t) = A|G(j)|sin(t + ) = tan
1
(
Im[G(j)]
Re[G(j)]
G in this case is the transfer function, with the complex
number j plugged in. The amplitude is affected by the
magnitude of the complex result, and the phase shift is the
angle. Both of these characteristics can be plotted with respect
to . When the amplication/attenuation factor is converted to
decibels, the plots become known as bode plots, and many
important system characteristics can be extrapolated from
them. The axis is plotted on a logarithmic scale, to account
for the large range of frequencies. The steady state response
of the system to any input frequency can be found by simply
locating the amplication and the phase shift on a bode plot
at that particular frequency and plugging the numbers into
AA448 - CONTROL IN AEROSPACE SYSTEMS LAB, WINTER 2013 5
the equation above. The DC gain can be extrapolated as
the frequency nears 0, and the resonance frequency of an
oscillating system can be found when the amplitude plot peaks
(the frequency where the system amplication of the signal is
highest). [1] A rough bode plot can also be constructed piece
by piece when the poles and zeros of a system are known, but
it is much more efcient to use a numerical tool. An example
of a bode plot is seen below in Fig. 5.
Fig. 5. Bode plot example [3]
VI. CONCLUSION
The contents of the course are concisely summarized in the
governing chart, and deal with how to approximately model
physical systems, solve those models, and then modify the
plant behavior without changing the system itself. Only a small
introduction to the eld of controls has been provided, yet this
course laid a solid foundation for further study of the subject.
REFERENCES
[1] Lum, Christopher. Control in Aerospace Systems - Lecture Notes Fall
2103 AA 447, University of Washington
[2] Oliveira, Mauricio. Lecture 14 - Stability MAE
280A, University of California San Diego
http://maecourses.ucsd.edu/mdeolive/mae280a/lecture14.pdf
[3] LJO. Control Tutorials for Matlab 27 Aug. 1996, Carnegie Mellon
Library, https://www.library.cmu.edu/ctms/ctms/matlab42/freq/freq.htm.

Vous aimerez peut-être aussi