Feedback Control

Massachusetts Institute of Technology Subject 2.017

 Components of Engineered Feedback Systems

• Plant: the system whose behavior is to be controlled.

Examples: vehicle attitude, temperature, chemical process,

business accounting, team and personal relationships, global
climate
• Actuator: systems which alter the behavior of the plant.

Examples: motor, heater, valve, law enforcement (!), pump, FET,

hydraulic ram, generator, US Mint
• Sensor: system which measures certain states of the plant.

Examples: thermometer, voltmeter, Geiger counter, opinion poll,

balance sheet, financial analyst
• Controller: translates sensor output into actuator input.
Examples: computer, analog device, human interface, committee

• Extreme variability in time scales:

– active noise cancellation requires ~100 kiloHertz sensing and actuation
– Social Security is assessed and corrected at ~3 nanoHertz (10 years)

Massachusetts Institute of Technology Subject 2.017

 Feedback fundamentally creates a new dynamics!

reference error input state measurement

6 Controller Plant Sensor
+ _

interest economic
Federal rates world indicators
6 analysts
+ _ Reserve economy

PWM duty shaft

cycle position
6 controller motor & load encoder
+ _

fuel

rate speed
6 cruise control engine & car tachometer
+ _

Massachusetts Institute of Technology Subject 2.017

 Basics in the
r e u y
Frequency Domain + _
C P

e=r–y

u = Ce = C(r-y)

y = Pu = PCe = PC(r-y) Æ (PC + 1)y = PCr Î y / r = PC / (PC + 1)

Similarly, e = r – y = r – PCe Æ (PC+1)e = r Î e / r = 1 / (PC + 1)

u = C(r-Pu) Æ (PC + 1)u = Cr Î u / r = C / (PC + 1)

Why can we do this? Convolution in time domain = Multiplication in freq. domain!

P must roll off at high frequencies – because no physical plant can respond to
input at arbitrarily high frequency. Same with controller C.

• Ideal case: e is a small fraction of r: e/r << 1, equivalent to y/r ~ 1

• This implies mag (PC + 1) >> 1 or mag (PC) >> 1.

• If plant P is given, then C has to be designed to make PC big.
• But mag (u / r) ~ mag(1 / P): HUGE when P gets small at high frequenciesÆ
excessive control action which will saturate or break actuators, excite
unmodelled plant behavior, etc.. Å issues of robustness

Massachusetts Institute of Technology Subject 2.017

 1

log |PC|
e/r = 1/(PC+1)
Performance

0
y/r = PC/(PC + 1) Robustness
Zc Zc
0
Z Z

Good tracking only possible at low frequencies Æ leads to a “formula” for design:

Make |PC| large at low frequencies, e/r ~ 0, y/r ~ 1;

Good regulation and tracking at low frequencies

Make |PC| small at high frequencies, e/r ~ 1, y/r ~ 0, u/r ~ C

Poor tracking at high frequencies, but reasonable control action

The frequency where |PC| = 1 is the crossover frequency Zc ;

Above this point, closed loop t.f. y/r = PC/(PC+1) drops off to zero.
So Zc is about the bandwidth of the closed-loop t.f.

Massachusetts Institute of Technology Subject 2.017

 LaPlace vs. Fourier XFM

Fourier Transform integrates x(t) e –jZt over the time range from
negative infinity to positive infinity

Laplace Transform integrates x(t) e-st over the time range from
zero to positive infinity

Result: X(jZ) can describe acausal systems, X(s) describes only causal ones!

Many important results of Fourier Transform carry over to LaPlace Transform:

L (x(t)) = X(s) (notation)
L (ax(t)) = a X(s) (linearity)
L (x(t) * y(t)) = X(s)Y(s) (convolution)
L (xt(t)) ÅÆ sX(s) (first time derivative)
L (xtt(t)) ÅÆ s2X(s) (second and higher time derivatives)
L (V x(t)dt) ÅÆ X(s) / s (time integral)
L (G(t)) = 1 (unit impulse)
L (1(t)) = 1/s (unit step)
Massachusetts Institute of Technology Subject 2.017
 LaPlace Transform and Stability

• For linear systems, stability of a system refers to whether

the impulse response has exponentially growing
components.
• No pre-determined input can stabilize an unstable system;

no pre-determined input can destabilize a stable system.

• Some examples you can work out:

L (e-Dt) = 1 / (s + D)

L (t e-Dt) = 1 / (s + D)2

L [ e-Dt sin(Zt) ] = Z / (s2 + 2Ds + D2 + Z2)

L [ Zd e-]Znt sin (Zdt) / (1-]2) ] = Zn2 / (s2 + 2]Zns + Z n2)

Major observation: stable signal ÍÎ roots of L denominator

have negative real parts: TRUE FOR ALL FIRST- AND
SECOND-ORDER SYSTEMS

Massachusetts Institute of Technology Subject 2.017

 Decoding the transfer function

Numerator polynomials are a snap:

(s + 2)/(s2+s+5) = s/(s2 + s + 5) + 2/(s2+s+5)

“input derivative plus two times the input, divided by the denominator”

For higher-order polynomials in the denominator: use partial fractions, e.g.,

(s+1)/(s+2)(s+3)(s+4) = -0.5/(s+2) + 2/(s+3) -1.5/(s+4) (all real poles)
(s+1)/s(s2+s+1) = -s/(s2+s+1) + 1/s (some complex poles)

Any high-order transfer function can always be broken down into a sum of transfer
functions with factored first- and second-order polynomials in the denominator.

stability ÅÆ the roots of the characteristic

equation have negative real part.

More details:

real negative root –D: the mode

decays with time constant 1/D

complex roots at -Zn] +/- jZd

the mode decays with frequency Zd

and exponential envelope having

time constant ]Zn

Massachusetts Institute of Technology Subject 2.017

 Example with a double integrator: e.g., a

motor or dynamic positioning

System is mxtt(t) = u(t) where:

m is mass

xtt(t) is double time derivative of position

u(t) is control action; thrust

Let a Control law be: u = - kp x (Proportional Control: P)

Closed-loop system dynamics become mxtt + kpx = 0
Response to an initial condition is undamped oscillations at frequency Zn = sqrt(kp/m)

P = 1/ms2
C = kp r e
C = kp P = 1 / ms2
PC = kp/ms2 Æ + _
e/r = 1/(PC + 1)

= ms2 / (ms2 + kp)

Tracking error is small when s is small; large when s is large, as desired.

BUT characteristic equation ms2 + kp = 0 has two imaginary poles – undamped!

Massachusetts Institute of Technology Subject 2.017

 Try the control law u = -kpx – kdxt (Proportional +
Derivative: PD)
Closed-loop system dynamics become mxtt + kdxt+ kpx = 0
Recall for a second-order underdamped oscillator:
0 < kd < 2 sqrt(kp/m)
Zn = sqrt(kp/m) (undamped natural frequency)
] = kd / 2 sqrt(kpm) (damping ratio)
Zd = Zn sqrt(1-]2) (damped natural frequency)

Response to an initial condition is either:

• Damped oscillations at frequency Zd = sqrt(1-]2)Zn,

inside an exponential envelope with time constant 1/]Zn
OR
• Sum of two decaying exponentials (overdamped case)

Massachusetts Institute of Technology Subject 2.017

 Consider a constant disturbance: mxtt + kdxt + kpx = F;
System will settle at x = F/kp; this is a steady-state error!
But kp cannot be increased arbitrarily – natural frequency
will be too high and too much control action

Try the control law u = -kpx – kdxt – ki V x dt

(Proportional + Derivative + Integral: PID)
Closed-loop system dynamics become

mxtt + kdxt+ kpx + ki V x dt = F

If the system is stable (ms3 + kds2 + kps + ki = 0 has roots

with negative real part), then differentiate:
mxttt + kdxtt + kpxt + kix = 0 Î settles to x = 0!

Massachusetts Institute of Technology Subject 2.017

 The PID

C = kp + kds + ki/s
= (kps + kds2 + ki) / s

High-frequency response is ~kds; increases with frequency

and disobeys our prior rule about infinite power. High
frequency errors will lead to very large control action!

Sensor noise solutions:

• use a very clean and high-res. sensor for x, which can
be easily differentiated numerically, e.g., motor encoder
• use a sensor that measures xt directly, e.g., tachometer
• filter the measurement. For a low-pass, we would get
Cf = [ (kps + kds2 + ki) / s ] X [ O/ (s+O) ]
= O (kps + kds2 + ki) / s (s+O)

But combine with a double integrator plant P = m/s2

PC = m(kps + kds2 + ki) / s3, which does go to zero at high

frequencies, as desired Æ the system does have a real
bandwidth, which can be tuned.

Massachusetts Institute of Technology Subject 2.017