utomatic control is a “miracle” tech- mechanical systems by block diagrams

nology that exploits feedback to Because automatic control is such an with forces, positions, velocities, and ac-
improve the performance of a tremendous intricate subject it is extremely easy for celerations as inputs and outputs and with
range of technological systems, from the students to miss the forest for the trees. the blocks representing kinematics and
steam engine to the space station. Watt’s This guide is intended to be of help to both dynamics.
governor tamed the steam engine and undergraduate and graduate students. Un- 3. Determine the equilibrium points
made the Industrial Revolution possible. dergraduates may skim this guide at the and linearize. An equilibrium point is a
However, feedback is used by everyone: beginning of a first course on control to state in which the system would remain if
if the shower water is too hot, open up the get arough roadmap of the subject and ter- it were unperturbed by external distur-
cold water faucet. Today, feedback con- minology and later refer to it periodically bances. An equilibrium point can be un-
trol is used in radios (amplifiers), CD during the course. At the end of the stable (an egg standing on its end),
players (laser tracking), automobiles course, these notes can be useful for re- neutrally stable (a mass connected to a
(cruise control, engines, and suspen- viewing the course while studying for the spring), or stable (a book lying on a table).
sions), flight control (autopilots and sta- final exam. For graduate students em- For a system under feedback control, an
bility augmentation), spacecraft (attitude barking on a course in modern control,
equilibrium point is called an operating
control and guidance), machine tools, ro- these notes can provide a quick review at
point.
bots, power plants, materials processing, the beginning of the course to help place
and many other applications. In many their prior knowledge in perspective. Real systems are nonlinear. How-
cases automatic control is an enabling ever, a linearized model can be used to
technology since the system cannot oper- approximate a nonlinear system near an
ate without it. Feedback is also used to 1. Feedback is pervasive. The inter- equilibrium point of the nonlinear sys-
regulate virtually every system in the hu- action of any pair of systems almost al- tem by a procedure called linearization.
man body and is constantly at work in eco- ways involves feedback. System l reacts The resulting linear system has an equi-
logical systems (but let’s not go too far to System 2 and vice versa. It is cascade librium point at zero that corresponds to
afield here). At first sight, control engi- (one-way interaction) that is the excep- the equilibrium point of the nonlinear
neering looks simple and straightforward. tional case. Newton’s third law is a state- system. While linearized models are only
It is not. While automatic control is apow- ment of feedback: The force applied to A an approximation to the nonlinear sys-
erful technology, the subject is amazingly by B is counteracted by the force applied tem, they are convenient to analyze and
subtle and intricate in both theory and to B by A. When two electrical circuits are they give considerable insight into the
practice. This subtlety is mainly because wired together, each affects the other, and behavior of the nonlinear system near the
changes cannot be effected instantane- Kirchoff‘s laws determine what that inter- equilibrium point. For example, if the
ously in a dynamical system-when the action is. zero equilibrium point of the linear sys-
shower is too hot to the touch, it is already 2. Block diagrams are not circuit tem is stable, then the equilibrium of the
too late to shut down the hot water. An diagrams. Block diagrams are helpful for nonlinear system is locally stable. Un-
otherwise correct control decision applied analyzing feedback. However, block dia- less stated otherwise, we will consider
at the wrong time could result in catastro- grams should not be confused with circuit only the approximate linearized model of
phe. Accounting for this and numerous diagrams. It is useful to be able to translate the system.
other effects is what control engineering circuit diagrams into block diagrams. 4. Check stabilityfirst. Once a linear-
is about. This translation requires representing im- ized model has been obtained, stability
pedances and admittances as systems with must be checked. Stability can be asked of
The author is with the Aerospace Engi- inputs and outputs that are voltages and any system, whether the system is uncon-
neering Department, The University of currents. Remember that the lines con- trolled (with the control turned off) or the
Michigan, A n n Aubor, MI 48109, necting the boxes are not wires. Analo- system is in closed-loop operation (with
dsbaero@umich.edu. gously, it is useful to be able to represent the control determined by feedback). If

96 IEEE Control Systems

Authorized licensed use limited to: MIT Libraries. Downloaded on January 14,2021 at 01:04:07 UTC from IEEE Xplore. Restrictions apply.
the system is represented by a transfer tain frequencies. For convenience we re- many times counterclockwise as there are
function, then the roots of the denomina- fer to these plots as the “frequency re- unstable (open right-half plane) poles in
tor polynomial (called the poles of the sponse” even when the system is not stable. the loop transfer function, whether these
system) determine whether the equilib- 6. Remember the loop transfer func- unstable poles arise from the plant or from
rium or operating point is stable. Stabil- the controller. Note that the Nyquist plot
tion. Breaking the loop of a closed-loop
ity can be tested explicitly by computing can only encircle the critical point -l+jO if
system reveals the loop transfer func-
the roots of this polynomial or implicitly the loop gain is greater than unity in some
tion, which is the product of all of the
by using the Routh criterion. If the sys- frequency range. Thus stabilization im-
transfer functions in the loop, including
tem is in state space form, then the eigen- poses a minimal requirement on the loop
the plant, the sensor, the controller, and
values of the dynamics matrix can be gain in certain frequency ranges. This re-
the actuator. Always be aware of whether
computed as an explicit test of stability. quirement thus imposes a constraint on
the loop transfer function of a closed-loop
Stability can also be tested by forming the the gain of the controller that depends
system is stable or not, and be sure to note
characteristic polynomial of the dynamics upon the gain of the plant.
whether the closed-loop system involves
matrix and applying the Routh criterion. The root locus procedure shows the
positive or negative feedback. The gain
However, this approach is inconvenient if location of the closed-loop poles for each
and phase of the loop transfer function at a
the system is of high order. value of a constant feedback gain. The
given frequency are called the loop gain
closed-loop poles are located near the
and loop phase, respectively, and these
5. Stable systems have a frequency poles of the loop transfer function for
quantities are defined whether or not the
response. Much of control-system small values of the feedback gain, and, as
loop transferfunction is stable. Note that
analysis involves the frequency re- the feedback gain goes to infinity, con-
if the loop transfer function is stable then
sponse of linear systems. The meaning verge to the locations of the loop transfer
it does not necessarily follow that the cor-
of the frequency response can be under- function zeros and move toward infinity
responding closed-loop system is stable,
stood by keeping in mind the funda- along asymptotes. Hence a nonminimum
and that if the loop transfer function is un-
mental theorem of linear systems: phase system, that is, a system with at
stable then it does not necessarily follow
Suppose that a sinusoidal (or harmonic) least one zero in the open right half plane,
that the corresponding closed-loop sys-
input (such as a forcing) with frequency can be destabilized by large constant
tem is unstable. Stability of a closed-loop
w is applied to a stable linear system G feedback gains. Furthermore, it can be
system can be tested directly by applying
(s). Then the output of the system ap- seen from the form of the asymptotes that
the Routh criterion to the closed-loop sys-
proaches a sinusoidal motion whose fre- tem or by computing the poles of the a system that has relative degree (pole-
quency is the same as the frequency of zero excess) of at least three can also be
closed-loop transfer function. In the case
the input. This limiting sinusoidal mo- of negative feedback, the closed-loop destabilized by large constant feedback
tion is called the harmonic steady- gains.
transfer function involves the sensitivity
state response. Note, however, that the 7. After stability, performance is
function which is one over one plus the
output of the system does not have a loop transferfunction. everything. Once stability is settled, per-
limit in the usual mathematical sense formance is everything. In fact, for plants
since the harmonic steady-state re- While the characteristics of the loop that are open-loop stable, Performance is
sponse does not approach a constant transfer function (such as its frequency re- the only reason for using feedback con-
value. The transient behavior of the sys- sponse and its poles and zeros) may be trol. Basic criteria for performance in-
tem before its response reaches har- well understood, the closed-loop system clude the ability to track or reject signals
monic steady state depends on the poles is relatively complicated. Thus it is much such as steps, ramps, sinusoids, and noise
and zeros of the system as well as on the more convenient to analyze stability indi- (random signals). Good performance
initial conditions of the internal states. rectly in terms of the loop transfer func- generally requires high loop gain which
The ratio of the amplitude of the har- tion. The Nyquist criterion and root corresponds to small gain of the sensitiv-
monic steady-state response to the am- locus procedure allow you to do just that. ity function and thus tracking error reduc-
plitude of the sinusoidal input is equal to While it is intuitively clear from the tion o r d i s t u r b a n c e attenuation.
the absolute value or gain of the transfer form of the sensitivity function that the Unfortunately, Bode’s integral theorem
function evaluated at the input fre- frequency response of the loop transfer tells us that for any real control system the
quency w, that is, 1/2G uw)l/2, while function of a stable system must never be sensitivity function cannot have gain less
the phase shift or phase of the harmonic equal to -1 if the closed-loop system is to than unity at all frequencies.
steady-state response relative to the be stable, the Nyquist criterion shows 8. Perfect performance is asymp-
phase of the input is given by the phase precisely how the frequency response of totically possible. In the extreme case,
angle of the transfer function evaluated at the loop transfer function must avoid this asymptotically perfect tracking or distur-
the input frequency, that is, tan-’(Im G value. Specifically, the Nyquist criterion bance rejection can be obtained by means
(iw)/Re G (iw)). Bode plots show the says that the closed-loop system with of infinite loop gain and thus zero sensi-
gain and phase of G (jw) for each fre- negative feedback is stable if and only if tivity at the disturbance frequencies. This
quency w in a range of frequencies. Al- the polar plot of the loop transfer function, is the case when an integral controller is
though the fundamental theorem of with its argument following a contour that used to provide infinite gain at DC (zero
linear systems does not apply to neutrally includes the imaginary axis, avoids poles frequency) and thus asymptotically track
stable and unstable systems, Bode plots of the loop transfer function on the imagi- a step command or reject a step distur-
can be drawn for these systems as well, nary axis, and encloses the right half bance. The same idea can be used to as-
although the gain may be infinite at cer- plane, encircles the critical point -l+jO as ymptotically track or reject a sinusoidal

August 1997 97

Authorized licensed use limited to: MIT Libraries. Downloaded on January 14,2021 at 01:04:07 UTC from IEEE Xplore. Restrictions apply.
signal by using infinite controller gain at function is 180 degrees, which quantifies improve performance in a frequency band
the disturbance frequency (which we are robust stability for a pure gain perturba- where the plant is well known.
assuming is known). Since the controller tion of the loop transfer function, while 14. High controller gain has lots of
incorporates a model of the command or the phase margin quantifies robust sta- benefits. If the controller gain is high and
disturbance, its operation is a special case bility for frequencies at which the loop the plant gain is not too small, then the
of the internal model principle. Don’t gain is unity (these are the crossover fre- loop gain will be high, which implies
forget that closed-loop stability must still quencies), which quantifies robust stabil- small sensitivity and thus good tracking or
be checked for this infinite-gain controller. ity for a pure phase perturbation of the good disturbance rejection. Also, assum-
9. Assure nominal stability. If the loop transfer function. The Nyquist crite- ing that nominal stability holds, high loop
system is uncertain, as least assure nomi- rion and the root locus procedure can both gain means large distance from the
nal stability. Everything that has been said be used to determine gain and phase mar- Nyquist plot of the loop transfer function
so far with regard to stability and perform- gins. Note that a closed-loop system with to the critical point -l+jO and thus some
ance is based upon the assumption that the nonminimum phase loop transfer func- measure of robust stability. Thus it would
linearized model of the plant is an accu- tion has limited gain margin (this was seem that high gain gives both good per-
rate representation of the system near the seen from the root locus procedure), while formance and robust stability. However,
equilibrium. While all real systems are a closed-loop system with loop gain there are (at least) three hidden catches to
nonlinear, it is convenient to view the lin- greater than unity at some frequency has having the best of both worlds, namely,
ear approximation itself as possessing limited phase margin (this follows from rolloff, saturation, and noise.
modeling uncertainty, so that the model the Nyquist criterion and Lesson 15 be- 15. Practice safe rolloff. Don’t forget
we have been discussing is merely a low). Never forget that the distance from (and we can’t stress this enough) that high
nominal model of the plant. All of the sta- the Nyquist plot to the critical point -l+jO loop gain is useless if the number of
bility tests, whether direct or indirect, can is irrelevant if the number of encir- Nyquist encirclements is wrong for nomi-
be applied to the nominal model to guar- clements is wrong, that is, if nominal sta- nal stability. Thus, as already noted, indis-
antee nominal stability. bility does not hold. That is why criminate use of high controller gain is not
indiscriminate use of large controller recommended. If nominal stability does
10. What guarantees robust stabil- gains is not a viable control strategy. hold, then there is still the problem of rol-
ity? Robust stability refers to stability loff, since the gain of a real system, and
for all linear models in the range of the 12. Always conserve phase. With the
critical point sitting at -l+jO, the Nyquist thus the loop gain, always goes to zero as-
modeling uncertainty. Robust stability ymptotically at high frequency. As the
can be guaranteed in principle by apply- criterion shows that any closed-loop sys-
loop gain drops below unity at the cross-
ing any stability test to all possible models tem is never more than 180 degrees from
over frequency, gain and phase margins
of the uncertain system. Unfortunately, instability at every crossover frequency.
must be maintained to provide adequate
this is hard work. However, the Nyquist And 180 degrees is not a whole lot, espe-
distance from the critical point -l+jO in
criterion can be used to determine robust cially at high frequencies where the plant
is more difficult to accurately model and order to assure robust stability against
stability with a frequency domain un- modeling uncertainty. In general, as the
certainty model, which measures the identify and thus the loop phase is more
likely to be uncertain. Every degree of frequency increases you will have to roll
level of modeling uncertainty at each fre- off the loop gain more quickly than the
quency. In general, uncertainty tends to be loop phase at crossover frequencies is pre-
plant uncertainty grows. Also remember
greater at higher frequencies, and loop cious and must be maintained by careful
engineering. This issue especially arises that achieving good rolloff isn’t as easy as
gain is generally known better than loop adding poles to the controller to roll off
phase. in digital control where analog-to-digital
the loop gain since, as the loop gain rolls
(sampling) and digital-to-analog (zero-
11. The Nyquist criterion can deter- order-hold) devices can cause phase lag.
off, the loop phase lags (that is, the loop
mine robust stability. The number of en- phase decreases) with 90 degrees of lag
circlements of the critical point -l+jO by 13. Beware of lightly damped poles. incurred at high frequency for each pole.
the Nyquist plot of the loop transfer func- Don’t forget that every lightly damped Hence good rolloff requires that the loop
tion is the crucial frequency domain test pair of complex conjugate poles in the gain decrease adequately without accu-
for nominal stability. Once the number of plant (or controller) entails high loop mulating excessive loop-phase lag. An-
encirclements is correct for nominal sta- gain near the resonance frequency as other Bode integral theorem shows that
bility, the distance from the Nyquist plot well as a whopping 180 degrees of phase most of the loop-phase lag is due to the
of the loop transfer function to the critical shift over a small frequency band. A rolloff rate at the crossover frequency
point -l+jO determines robust stability in notch in the controller can reduce the with steeper rolloff causing greater phase
terms of a frequency domain uncertainty loop gain at the resonance frequency to lag. Lead-lag compensators are useful
model. This distance is the reciprocal of help avoid an inadvertent change in the for shaping the loop gain and loop phase
the gain of the sensitivity. Thus, small number of Nyquist encirclements due to to achieve high gain and safe rolloff.
sensitivity at a given frequency corre- modeling uncertainty. However, this 16. Saturation can rob you. As if
sponds to large distance from the Nyquist technique will only work if you actually that’s not all the trouble high loop gain can
plot of the loop transfer function to the know the frequency of the resonance. get you into, high loop gain is useless if
critical point -l+jO and hence robust sta- Note that a notch in the controller entails the actuators cannot deliver the specified
bility at that frequency. The gain margin some performance loss in the frequency control signal. If the controller asks for
measures robust stability for frequencies band of the notch. An antinotch can be four Newtons and the actuator can deliver
at which the phase of the loop transfer used to increase the loop gain and thus only two Newtons, then you have a seri-

98 IEEE Control Systems

Authorized licensed use limited to: MIT Libraries. Downloaded on January 14,2021 at 01:04:07 UTC from IEEE Xplore. Restrictions apply.
ous problem. You can think of the Nyquist
plot as being “squashed down” due to this
saturation effect. Margins will be re-
duced, and, even worse, the Nyquist en-
circlement count can change and the
closed-loop system will be unstable. It is
extremely important to remember that the
inability of the actuators to deliver the
specified control signal is not just the fault
of the controller gain being too high, but
rather is also due to both the size of the
plant gain and the amplitude of the distur-
bance signal. If the disturbance signal has
large amplitude, then the actuator may
saturate and you will have no choice but to
reduce the gain of the controller and thus
sacrifice some of the performance you
want (and which was the reason for using
feedback control in the first place!).
Therefore, saturation can rob you of both
stability and performance. Although we
discussed saturation in terms of linear sta-
bility analysis, saturation is actually a spe-
cial kind of nonlinearity.
17. High gain amplifies noise. If your
high gain controller survives the rolloff
dragon and the saturation beast, it may yet
succumb to the noise monster. While an
integrator (“l/s”) tends to smooth and at-
tenuate noise, a differentiator (“s”) tends
to amplify noise. Every pole in a transfer
function is “like” an integrator, while
every zero is ‘‘like’’a differentiator. As the fer function of the actual system may be 20. Multi-loop control is nontrivial.
plant gain rolls off, you may wish to in- different from what you expect, and the Multi-loop control is much more chal-
clude zeros in your controller in order to closed-loop system may be unstable. If lenging than single-loop control. Every-
increase the loop gain for better perform- you know about the delay in advance, then thing that has been said so far applies to
ance while adding phase lead to the loop you may be able to counteract the effect of single-loop control. Multiple control
transfer function in order to increase your the delay through careful loop-shaping loops are needed whenever a plant has
phase margin for robust stability. Zeros design. However, although you can (at multiple sensors or actuators. In this case
will do both of these things quite nicely least in theory) design stabilizing control- the interaction of every feedback loop
for you. However, the resulting high con- lers in the presence of even large time de- with every other feedback loop must be
troller gain will now amplify noise in the lays, the closed-loop performance will accounted for. While many single-loop
measurements, and this amplification most likely be poor. Imagine trying to concepts hold in principle in the multi-
may outweigh the performance and stabil- make decisions using old information. loop case, the technicalities are much
ity benefits of the high loop gain and 19. “Respect the unstable.” As Gun- more involved. The performance benefits
loop-phase lead. This amplification of ter Stein discussed in his classic 1989 of multi-loop control, however, are often
noise was to be expected since the zeros, Bode lecture (watch the video!), control- far more than one would expect from a
after all, act like differentiators. ling unstable systems can be a dangerous collection of single-loop controllers.
18. Time delays can be deadly. A undertaking. Real unstable plants (except 21. Nonlinearities are always pres-
time delay can be deadly-think of the for rigid body motion) are always nonlin- ent. Almost everything that has been said
Trojan horse or the AIDS virus. A time de- ear, and the ability to stabilize them re- so far applies to linearized plant models.
lay in the feedback loop corresponds to a quires a minimal amount of actuator Real systems, however, have all kinds of
transfer function that has unit magnitude bandwidth and stroke. A disturbance may nonlinearities: deadband, backlash, Cou-
at all frequencies as well as phase lag that perturb the state of a nonlinear system out- lomb friction, hysteresis, quantization,
increases linearly with frequency. This side of its domain of attraction so that saturation, kinematic nonlinearities, and
phase lag will warp the Nyquist plot, espe- the actuators are unable to move the state many others. Thus a controller designed
cially at high frequencies. Thus, if a back to the equilibrium and thus recover for a linear plant model to satisfy per-
closed-loop system has an unmodeled from the disturbance. In addition, actuator formance specifications may perform
time delay in the loop, then the number of failure can lead to disaster and thus cannot poorly when applied to the actual plant!
Nyquist encirclements by the loop trans- be tolerated. The tradeoff here is between mathemati-

August 1997 99

Authorized licensed use limited to: MIT Libraries. Downloaded on January 14,2021 at 01:04:07 UTC from IEEE Xplore. Restrictions apply.
cal tractability of the linearized model and (differential equation) models and trans- automatic control will become the technol-
greater validity of a nonlinear model. fer function matrices. State space models ogy of the next millennium. You are defi-
22. People’s lives may be at stake. provide the means for designing control nitely in the right place at the right time!
Control-system engineers must account systems that are optimal with respect to
for all of these issues in designing and certain design criteria. Optimal control
analyzing control systems that work. encompasses LQG and H. control theory
They must also specify, design, analyze, where explicit formulas are used to syn- tory of automatic control in:
build, program, test, and maintain the thesize multivariable feedback control- 0. Mayr, The Origins of Feedback Control, MIT
electromechanical hardware, processors, lers. Robust control seeks controllers Press, Cambridge, 1970.
and software needed to implement control that provide robust stability and perform-
S. Bennett, A History of Control Engineering
systems. Real control systems must be ex- ance for uncertain plants. The computer 1800-1930,Peter Peregrinus Ltd.,London, 1979.
tremely reliable, especially if people’s implementation of these controllers is the
subject of digital control.While the jump S. Bennett, A History of Control Engineering
lives depend on them. These are challeng-
1930-1955,Peter Peregrinus Ltd., London, 1993.
ing and rewarding engineering tasks that from single-variable control to optimal
will keep lots of control-system engineers multivariable control is a major and im- To get a glimpse of the future and what hath Watt
busy for a long time to come. portant step, nonlinear control takes wrought, see the exciting book:
nonlinearities into account and shows K. Kelly, Out of Control, Addison-Wesley, Read-
how to design nonlinear controllers to ob- ing, 1994.
Now that you have completed the ba- tain improvements over linear control.
sic course in classical control. you are The next, and most exciting, leap is into
ready to enter the wonderland of modern the subject of adaptive control, where I would like to thank an anonymous re-
control. The linear control you have controllers learn and adapt in response to viewer and Steve Yurkovich for numer-
learned, which has been limited to changes in the plant and disturbances. ous helpful suggestions. I especially
single-input, single-output systems, can With amazing advances in sensing, ac- thank all of my students, graduate and un-
now be expanded to multiple-input, tuation, and processing as well as a better dergraduate, who surely taught me far
multiple-output plants. Multivariable understanding of learning and adaptation more than I taught them. This guide is
control is often studied with state space (for example, using neural computers). their legacy.

1997 IEEE Fellows

curity assessment of power systems, power system operation and control, and application of robust
control techniques to power systems. He is the author and co-author of several papers in his field. He
has also co-authored a textbook entitled Power System Tranisent Stability Assessment Using the Tran-
sient Energy Function Method. During 1993-1994 he was the Program Director of the Power Systems
Program at the U.S. National Science Foundation. He is a recipient of the 1985 Presidential Young In-
vestigator Award. In 1988, he received the NCR Faculty Award of Excellence. He also received the
1989 Iowa State University College of Engineering “Young Engineering Faculty Research Award.”

Yuan F. Zheng
The Ohio State University
For contributions to the development of mechanisms for coordination of multiple robots and manipu-
lams.
Yuan F. Zheng received M.S. and Ph.D. degrees in electrical engineering in 1980 and 1984, respec-
tively, from The Ohio State University in Columbus, OH. His B.S. degree was received in 1970 at
Tsinghua University in Beijing, China. From 1984 to 1989, he was with the Department o f Electrical
and Computer Engineering at Clemson University in Clemson, SC. He has been with the Ohio State
University since 1989, where is currently a professor and is chairman of the Department of Electrical
Engineering. His research interests include coordination of multiple robots for both manipulators and
walking machines, coordination of human arm and robot, and wavelet transform for image compres-
sion and surface digitizing. He received the Presidential Young Investigator Award in 1987 for his re-
search contributions. He is Vice-president for Technical Affairs o f the IEEE Robotics and Automation
Society, an associate editor of the IEEE Transactions on Robotics and Automation, on the Editorial
Board of Autonomous Robots, an associate editor of the International Journal of Intelligent Autorna-
tion and Soft Computing, and on the editorial board of the International Journal of Intelligent Control
and Systems. He has served on the program committees of many international conferences and will be
the Program Committee Chairman of the 1999 International Conference on Robotics and Automation.

100 IEEE Control Systems

Authorized licensed use limited to: MIT Libraries. Downloaded on January 14,2021 at 01:04:07 UTC from IEEE Xplore. Restrictions apply.