Professional Documents
Culture Documents
L2 Robust
L2 Robust
1. Introduction
Robust Control
2. Comparing two Systems
3. Robustness Measures
Karl Johan Åström and Anders Rantzer
4. Classic Loop Shaping
Automatic Control LTH
5. H∞ Loop Shaping
Lund University
6. Gain scheduling
January 20, 2021 7. Summary
Similar Open Loop Different Closed Loop Different Open Loop Similar Closed Loop
Open loop
100 Open loop
500
400
50
y
300
y
200
0 100
0 1 2 3 4 5
0
t 0 0.5 1 1.5 2
Closed loop t
3 Closed loop
2
1
1
y
0 0.5
−1
0 0.1 0.2 0.3 0.4 0.5
0
t 0 0.02 0.04 0.06 0.08 0.1
t
100 100
P1 (s) = , P2 (s) =
s+1 (s + 1)(1 + 0.025s)2 Systems and complementary sensitivity functions
Complementary sensitivity functions with unit feedback C = 1
100 100 100 100
100 1.616e5 P1 (s) = , T1 (s) = , P2 (s) = , T2 (s) =
T1 = , T2 = s+1 s + 101 s−1 s + 99
s + 101 (s + 83.9)(s2 − 2.90s + 1926s + 1926)
The closed loop systems are very different even if open loop step Closed loop systems are very similar even if open loop systems step
responses are very close! responses are very different!
1
Vinnicombe’s ν -gap Metric δ ν Return to the Motivating Examples
Open loop Open loop
100 500
400
300
50
y
200
100
0 0
0 1 2 3 4 5 0 0.5 1 1.5 2
t t
Closed loop Closed loop
3
1
2
y
0.5
0
−1 0
0 0.1 0.2 0.3 0.4 0.5 0 0.02 0.04 0.06 0.08 0.1
t t
Winding number constraint is satisfied and
pP1 (iω ) − P2 (iω )p 100 100
δ ν (P1 , P2 ) = sup p ∈ [ 0, 1] P1 (s) = , P2 (s) = , δ ν (P1 , P2 ) = 0.98
ω (1 + pP1 (iω )p2 )(1 + pP2 (iω )p2 ) s+1 (s + 1)(1 + 0.025s)2
if P1 can be transformed smoothly to P2 without the map of the Nyquist
100 100
curve flipps over the north pole. If not, the ν -gap metrix is defined to be 1. P1 (s) = , P2 (s) = , δ ν (P1 , P2 ) = 0.02
s+1 s−1
1. Introduction
2. Comparing two Systems
3. Robustness Measures
◮ Comparing time responses can be misleading! Process variations
Stability margins gm , φ m , Ms , Mt
◮ Frequency responses are better Delay margin
◮ Vinnicombe’s ν -gap metric is a good measure Small process variations
Large process variations
4. Classic Loop Shaping
5. H∞ Loop Shaping
6. Gain scheduling
7. Summary
r e u v x y
F Σ C Σ P Σ
◮ Broad descriptions
Classical - gain gm and phase margins φ m
Additive ∆, multiplicative δ and feedback uncertainty ∆fb
−1
Maximum sensitivities Ms and Mt
◮ Specific characterization Controller Process
Specific process parameters
1 P PCF
6 n n
Time delays - delay margin y G Gyd Gyr
5 6 5
= yn u = 1 +CPC 1 + PC
1
1 + PC
CF d
◮ Small variations u Gun Gud Gur
r − − r
◮ Large variations 1 + PC 1 + PC 1 + PC
◮ Command signal response can always be fixed by feedforward F
◮ First two columns suffice for robustness (Gang of Four)
◮ Physical interpretations
Gyd - response of the output y to load disturbance d
Gun - response of control signal u to measurement noise n
Nominal process
Perturbed process
P Σ P Σ
P = (N + ∆N )(N + ∆M )−1 , MM ∗ + NN ∗ = I
∆ δ Block diagram
Σ P
∆fb
∆M ∆N
u y
S M −1 N S
v
2
Stability Margins and Maximum Sensitivity A Difficulty with Gain and Phase Margins
Gain margin
Im PC(iω ) Im PC(iω )
Ms
gm ≥ −1
Ms − 1
1/Ms ω ms −1 −1/gm Re PC(iω )
Phase margin 1/Ms Re PC(iω )
φm
1 ωs
φ m ≥ 2 arcsin
2Ms
Constraints on both gain and phase margins can be replaced by one ◮ Necessary to specify both gm and φ m
constraint on Ms . ◮ Not sufficient to specify both gm and φ m (right figure)
◮ Ms = 2 guarantees gm ≥ 2 and φ m ≥ 30○ ◮ Ms can replace φ m and gm
◮ Ms = 1.6 guarantees gm ≥ 2.7 and φ m ≥ 36○ ◮ φ m and gm are widely used in industry difficulties when the Nyquist
◮ Ms = 1.4 guarantees gm ≥ 3.5 and φ m ≥ 42○ curve has warts
◮ Ms = 1 guarantees gm = ∞ and φ m ≥ 60○
Delay Margin - Time delay required to make system unstable Small Variations in Process Dynamics
The behavior of the closed loop system is captured by the Gang of Four,
how do they change when there are small errors in the process
2
pPC(iω )p
PC dT 1 dP dP
0
10
1 T = , = =S
1 + PC T 1 + PC P P
ImPC(iω )
−2
10
0
1 dS −PC dP dP
10
−2 −1
10
0
10
1
10
2
10
S= , = = −T
−1
1 + PC S 1 + PC P P
P dGyd dP
0
−2
Gyd = , =S
∠PC(iω )
−3
−360
1 + PC Gyd P
−4 C dGun dP
Gun = , = −T
−720
−4 −2 0 2
RePC(iω ) 10
−2 −1
10
ω
0
10
1
10
2
10
1 + PC Gun P
Notice that
Peaks in the loop transfer function PC are dangerous
◮ S+T =1
They are often caused by resonances
◮ S is small at low frequencies and S ( 1 for high frequencies
100 0.25
P (s) = , C(s) = , L = 0.3 ◮ T small at high frequencies T ( 1 at low frequencies
s(0.5s + 1)2 (s2 + 0.004s + 100) s
◮ Large sensitivities are particularly helpful for T and dGyd /dt, effect of
modeling errors significantly less than dP /P when pSp is small
1 + PC
B δ δ
A C ∆P
P Σ
− 1+PCPC
−C
∆P 1 + PC 1 1
pC∆P p < p1 + PCp, < = ≤ The small gain theorem gives the stability condition
P PC pT p Mt
1 + PC 1
◮ Large variations permitted when T is small pδ p < =
PC pT p
◮ Small variations when T is large, Mt = max pT (iω )p
Same result as obtained before!
◮ Mt = 1.4 and 2 permits modeling errors of 70% and 50%
3
Harry Nyquist 1889-1976 The Nyquist Plot
◮ Strongly intuitive Im L(iω )
◮ Stability and Robustness
From farm life in Nilsby Värmland to Bell Labs
Stability margins φ m , gm ,
−1 −1/gm Re L(iω )
sm = 1/Ms
Dreaming to be a teacher sm
Frequencies ω ms , ω gc , ω pc
◮ Emigrated 1907 φm
◮ Disturbance attenuation
◮ High school teacher 1912 Circles around −1, ω sc
◮ MS EE U North Dakota 1914 ◮ Process variations
◮ PhD Physics Yale 1917 Easy to represent in the Nyquist plot
Parameters sweep and level curves of pT (iω )p
◮ Bell Labs 1917
◮ Measurement noise not easily visible
Key contributions
◮ Command signal response
◮ Johnson-Nyquist noise
Level curves of complementary sensitivity function
◮ The Nyquist frequency 1932 ◮ Bode plot similar but easier to use for design because its wider
◮ Nyquist’s stability theorem frequency range
the large value which it has in the useful frequency band to zero dB or less at
10
Z ω gc [
0
10
to reduce the gain very rapidly. The more rapidly the feedback vanishes for −1
10 10
0 1
10
log ω
example, the narrower we need make the region in which active design attention is
required to prevent singing. ... −90
−135
But the analysis in Chapter XIV (Bode’s relations) shows that the phase shift is
−180
broadly proportional to the rate at which the gain changes. ... A phase margin of
−1 0 1
10 10 10
log ω
Requirements An Example
Translate requirements on tracking error and robustness to demands on
◮ Stablity and robustness
the Bode plot for the radial servo of a CD player.
Gain margin gm , phase margin φ m , maximum sensitivity Ms
p∆P p p1 + PCp
Stability for large process variations: < ,
pP p pPCp
◮ Load disturbance attenuation
Ycl (s) 1
=
Yol (s) 1 + PC
◮ Can be visualized in Hall and Nichols charts
◮ Measurement Noise
U (s) C
− =
N (s) 1 + PC
◮ Command signal following (system with error feedback)
PC
T = can be visualized in Hall and Nichols charts
1 + PC From Nakajima et al Compact Disc Technology, Ohmsha 1992, page 134
Fix shape with feedforward F
Major disturbance caused by eccentricity about 70 µm, tracking
How are these quantities represented in loop shaping when we typically requirements 0.1 µm, requires gain of 700, the RPM varies because of
explore Bode, Nyquist or Nichols plots? constant velocity read out (1.2-1.4 m/s) around 10 Hz.
4
Nathaniel B Nichols 1914 - 1997 Hall and Nichols Chart
◮ BS Chemistry Central Michigan University 1936
3
◮ MS Physics University of Michigan 1937
4
◮ Taylor Instruments 1937-50 2
log pL(iω )p
2
◮ MIT Rad Lab 1942-46
Im L(iω )
◮ Raytheon 1951-55 0 1
◮ Taylor Instruments 1957-63 −2
◮ Aerospace Corporation 1963 - 67 0
Red suspenders! −4
−1
Accomplishments −5 0 5 −4 −3 −2 −1 0
Re L(iω ) arg L(iω ) [rad]
◮ Ziegler-Nichols tuning rules
◮ Servomechanism theory in James Nichols Phillips Theory of Hall is a Nyquist plot with level curves of gain and phase for
Servomechanisms 1947 the complementary sensitivity function T . Nichols=log Hall.
◮ The Nichols chart Both make is possible to judge T from a plot of PC
◮ Rufus Oldenburger Medal 1969 Conformality of gain and phase curves depend on scales
The Nichols chart covers a wide frequency range
◮ Richard E Bellman Control Heritage Award 1980
The Robustness Valley Re L(iω ) = −1/2 dashed
y = Pu, u = C (r − y )
The cost of feedback is that a controller which has high gain over a wide
10
pGun (iω )p
frequency range is required. The high controller gain means that small
measurement errors may generate large control signals that could cause
0
10
saturations in the system. If there is measurement noise the high gain may
also cause control actions that may wear out the actuators.
−1
10
−1 0 1 2
10 10 10 10
The transfer function from measurement noise n to the control signal for a
closed loop system is
pP (iω )p pT (iω )p
0
10
C T
−Gun = =
1 + PC P
−2
10
ω
max pGun (iω )p eller ppGun pp2 .
Notice that controller gain is high way beyond
the gain crossover frequency marked with the red dots
5
QFT Design Cost of Feedback - Robust stability for 2DOF
log(P )
“Quantitative Feedback Theory” (QFT) is a good choice when the main ω
problem is model uncertainty.
Often gives a good intuitive understanding of the design problem
log(Pm )
Carpentry in the Nichols chart log( PPm )
Its main strength is for SISO design, but MIMO and nonlinear extensions
exist. Will only discuss SISO design here
P Cff
p∆p < , ∀ω
Pm Cfb
Regions where gain is increased and where feedback is used require
low model uncertainty
Templates An Observation
ka
P (s) = , k ∈ [ 1, 5] , a ∈ [ 1, 4]
s(s + a) Note: The set {P (iω )C(iω ))}P ∈P has the same shape as the template
{P (iω )}P∈P ! Not true in Nyquist.
Template is moved by (180/pi*angle(C),20*log10(abs(C)))
Nyquist templates for w=[1 3 10] Nichols templates for w=[1 3 10]
1
0.5
10
For each frequency ω k :
0
−0.5
0
◮ Move the template around in the Nichols chart
◮ Mark all positions of the nominal point for which the entire template
−1
−1.5 −10
−3
−30
−3.5 “Horowitz Bounds”
−4 −40
−4 −3 −2 −1 0 1 −270 −180 −90 0
d n
1. Introduction
y
2. Comparing two Systems ❡
❄ ✲ P ❡
❄ ✲
3. Robustness Measures
4. Classic Loop Shaping
5. Loop Shaping for Specific Process Variations u
✛ K ✛
6. H∞ Loop Shaping
7. Gain scheduling
8. Summary y I $ n
5 6 5 6 5 6
(I + PC)−1 I P
#
=
u C d
6
Use weighting matrices!
ds ns
ds ns
❞ ✲ W1
❄ ❞
❄
❞ ✲ W1 ❞
P W2 ✲
❄ W2
❄ ✲ ys
P
ys
us
us ✛ K ✛
✛ K ✛
1) Choose W1 and W2 and absorb them into the nominal plant P to get
the shaped plant Ps = W2 PW1 .
20
W2 ( i ω ) P ( i ω ) W1 ( i ω )
Magnitude
10
−10
−20
(us , ys ). If the gain is large, the return to Step 1.
Measurement errors
−30
−40
3) The final controller is C = W1 C∞ W2 .
−50
−60
w0 w1
Frequency
−70 0 1
1. Introduction
2. Comparing two Systems
3. Robustness Measures
4. Classic Loop Shaping
5. Loop Shaping for Specific Process Variations
6. H∞ Loop Shaping
7. Gain scheduling
What is it?
How to find schedules?
Applications
8. Summary
Controller
parameters Gain
schedule
Operating
condition
◮ Select scheduling variables
Command
signal Control ◮ Make control design for different operating conditions
signal
Controller Process Output ◮ Use automatic tuning
◮ Transformations
7
Schedule on Process Variable Schedule on External Variable
(a) Output c
cr
1.0
q = 0.5
0.5 q = 0.9
q = 1.1
q =2
0.0
0 5 10 15 20
Time
(b) Control signal cin
q = 0.5
1.5 q = 0.9
The process can be modeled by an FOTD system where both the time 1.0
constant and the lag depend on the flow rate. The system will be controlled q = 1.1
0.5
by a sampled controller where the sampling rate depends on the flow rate - q=2
sample per meter and not per time. Similar ideas are used in control of 0.0
rolling mills where the time delay also depends on machine speed. 0 5 10 15 20
Time
q = θ˙ Pitch stick
Gear
VIAS
α Position
θ
V Filter A/D K DSE
Σ
H M
H
−
Nz δe T1s
K SG
1+ T1s
Operating conditions Σ D/A Σ
80 Acceleration
Filter A/D D/A Σ
60
Altitude (x1000 ft)
To servos
M Filter
Pitch rate
40
Σ Σ
1
3 4 Filter A/D K Q1 K NZ
1+ T3 s
20
M H
1 2 T2 s
K QD
0 1+ T2 s
0 0.4 0.8 1.2 1.6 2.0 2.4
Mach number
H M VIAS
8
Schedule of KQ with Respect to Indicated Airspeed (IAS) and Igelsta Power Plant for District Heating
Height (H)
Igelsta 120 MW co-generation plant in Södertälje. Heat exchanger with
nonlinear valve.
K QD IAS K QD H
0.5
0.5
0 0
0 1000 0 10 20
V IAS (km/h) H (km)
An ordinary PID controller was replace with a PID controller having gain
scheduling. Operating regions were set manually. The schedule was
determined by relay auto-tuning.
Valve position K Ti Td
0.00-0.15 1.7 95 23
Modified system 0.15-0.22 2.0 89 22
0.22-0.35 2.9 82 21
0.35-1.00 4.4 68 17
Original Modified
0.5
0
y
−0.5
−1
0 5 10 15 20 25 30
t
9
Summary
10