Design of Multi-loop control systems

Consider a single loop system as shown in Fig.1.

Fig.1

Suppose controller is fixed, substantial changes in K p invariably lead to deteriorate the

control system response (see Figure 2).

Fig.2

Now consider a 2 × 2 control system in Fig. 3

Fig.3

Let us consider open loop 1 (Fig.4)

1
 Fig.4

One may notice that there are two transmission paths from m1 to x1 .let us define a
relative gain from m1 to x1 .as:
Gain m1 − x1 , loop 2 open
λ=
Gain m1 − x1 , loop 2 close
When λ = 0.5 , the responses of x1 to a unit step input at m1 is shown in Fig. 5

Fig.5

In this case, when loop 2 is closed, the open loop gain of m1 − x1 becomes doubled.
The increase in the loop gain results in more oscillation in the closed loop response as
shown.
On the other hand, when λ = 2 , the open loop and closed loop responses are also

2
given in Fig. 6.

In this latter case, the open-loop gain decrease when loop 2 is switched from open to
close. As a result, the close of loop 1 leads the system to a more sluggish response to
the r1 input.
The increase or decrease of the loop gain is a result of closing loop2, and , hence, is
considered loop interaction. From the above example, λ is a measure of such
interaction and is named as relative gain of loop 1. You may also find the other
relative gain for loop2. But, in this case, the two relative gains will be equal.

Algebraic Properties of the RGA

1. ∑g
j
ij g ji = ∑ λ
j
ij = 1, ∀i

2. P1Λ{G} P2 = Λ{P1G P2}, P1 and P2 are two permutation matrices.

3. Λ{G} = Λ{S1G S2}, S1 and S2 are two diagonal matrices.
4. If transfer matrix, G, is diagonal or triangular, then: Λ{G}=I.
[Proof]:
 x 0 0...0 
 x x 0...0 
Let, G =  
.... 
 
 x x x... x 
Then,

3
  x 0 0...0 
 x x 0...0 
G −1 =G = 
.... 
 
 x x x... x 
Thus, g ij g ji = 0 = λ ij, ∀j ≠ i

and, g ii g ii = 1 = λ ii , ∀j = i

So, Λ =I

dg ji dg ij
5. = −λ ij
g ji g ij

adj[ A] ( −1) i + j det[G ij ]

g ij = =
det[G ] det[G ]
d det[G ]
( −1) i + j det[G ij ]
dg ji dg ij det[G ij ]2
=− 2
= 2
= − g 2ji
dg ij det[G ( g ij )] det[G ( g ij )]

dg ji dg ij dg ij
= − g ji dg ij = − g ji g ij = −λ ij
g ji g ij g ij

dλ dg ij dλ λ ij − 1 dg ij
= (1 − λ ij ) =
ij ij
5. , and
λ ij g ij λ ij λ ij g ij

[Proof];

λ ij = g ij g ji ⇒ dλ ij = dg ij g ji + g ij dg ji

dλ dg ij g ji + g ij dg ji dg ij dg ji dg ij
⇒ = = +( ) = (1 − λ ij )
ij

λ ij g ij g ji g ij g ji g ij

or,

dλ  dg ij  dg ji 1 dg ij  λ ij − 1 dg ij
⇒ = + = (1 − =
ij
 g  g
) 
λ ij  ij  ji λ ij g ij  λ ij  g ij

4
 RGA-implications:

1. Pairing loops on λij values that are positive and close to 1.

2. Reasonable Pairings: 0.5 < λij< 4.0

3. Pairing on negative λij values results in at least one of the following;

a. Closed loop system is unstable,
b. Loop with negative λij is unstable,
c. Closed loop system becomes unstable if loop with negative is λij turned off.

4. Plants with large RGA-elements are always ill-conditioned. (i.e., a plant with a
large γ(G) may have small RGA-elements)

5. Plants with large RGA-elements around the crossover frequency are fundamentally
difficult to control because of sensitivity to input uncertainties.
-----
decouplers or other inverse-based controllers should not be used for plants
with large RGA-elements.

6. Large RGA-element implies sensitivity to element-by-element uncertainty.

7. If the sign of RGA-element changes from s=0 to s= ∞ , then there is a RHP-zero

in G or in some subsystem of G.

8. The RGA-number can be used to measure diagonal dominance:

RGA-number = || Λ(G)-I ||min.
For decentralized control,, pairings with RGA-number at crossover frequency
close to one is preferred.

9. For integrity of whole plant, we should avoid input-output pairing on negative

RGA-element.

10. For stability, pairing on an RGA-number close to zero at crossover frequency is

preferred.

5
 The Relative Disturbance Gain (RDG)

Ref: Galen Stanley, Maria Marino-Galarraga, and T. J. McAvoy, Shortcut

Operability Analysis. 1. The relative disturbance gain, I&EC, Process Des. Dev.
1985,24, 1181-1188

The use of RDG:

1. To decide if interaction resulting from a disturbance is favorable or unfavorable.
2. To decide whether or not decoupling should be used and what type of decoupling
structure is best.

y1 = k11m1 + k12 m2 + k F 1d
y2 = k21m1 + k22 m2 + k F 2 d

 ∂m1  k
 ∂d  = − F1
  y1 , m2 k11

 ∂m1  is derived when both y1 and y2 are held still:

 ∂d 
  y1 , y2

y1 = k11m1 + k12 m2 + k F 1d = 0 (2)

y2 = k21m1 + k22 m2 + k F 2 d = 0
so that:

1
m2 = [ −k21m1 − k F 2 d ] (3)
k22
Substitute Eq.(3) into E.(2), we have:

 k12 k21   k12 k F 2 

k
 11 − m +
 1  F1k − d = 0
 k 22   k 22 

Thus,
k12 k F 2
−kF1 + (4)
 ∂m1  k22 k k − k F 1k22
 ∂d  = = 12 F 2
  y1 , y2 k k k11k22 − k12 k21
k11 − 12 21
k22

6
So,

 ∂m1 
 ∂d 
  y1 , y2 k k k − k F 1k22 k11 k22 k12 k F 2 − k F 1k22 (5)
β1 = = − 11 × 12 F 2 = ×
 ∂m1  k F 1 k11k22 − k12 k21 k F 1k22 k11k22 − k12 k21
 ∂d 
  y1 , m2
k12 k F 2 − k F 1k22 k11k22  k k 
=- × = 1 − 12 F 2  λ
k F 1k22 k11k22 − k12 k21  k F 1k22 

Similarly, we have:

 ∂m1 
 ∂d 
  y1 , y2  k21k F 1 
β1 = = 1 − λ
 ∂m1   k k
F 2 11 
 ∂d 
  y1 , m2

k F 2 k12 β λ − β1 k F 2 λ − β1 k22
= 1− 1 = ⇒ = ×
k F 1k22 λ λ kF1 λ k12

k F 1 λ − β 2 k11
Similarly, ⇒ = ×
kF 2 λ k21

So, kF 2 kF 1  λ − β 2 k11   λ − β1 k22 

× =1=  × × × 
kF1 kF 2  λ k21   λ k12 

 λ − β2   λ − β1  k11 k22 1
⇒  λ   λ  = k k = 1− λ
   21 12

or,

( β1 − λ )( β 2 − λ ) = λ (λ − 1)
λ (λ − 1) λ (λ − 1) ( β − 1) λ 1 − β1
( β2 − λ ) = ⇒ β2 = +λ = 1 = λ
( β1 − λ ) ( β1 − λ ) ( β1 − λ ) λ − β1
It can be shown that:

7
 Multi-loop e1 area
∝ β1 and
SISO idealy decoupled e1 area
Multi-loop e 2 area
∝ β2
SISO idealy decoupled e 2 area

e1 F2 G12 Gc 2 − F1 (1 + G22 Gc 2 ) (6)

=
d (1 + G11Gc1 )(1 + G22 Gc 2 ) − Gc 2 Gc1G12 G21

If d is a unit step, then the area under e1 curve is given as:

k k k k
∞ k F 2 k12 c 2 − F 1 22 c 2
τ R2 τ R2 k F 2 k12 − k F 1k22
∫0 e1dt = lim
s →0
e1 ( s) =
kc1k11 kc 2 k22 kc 2 kc1
=
kc1k11k22  k12 k21 
− k k 1−
τ R1 τ R 2 τ R 2 τ R1 12 21 τ R1  k11k22 
τ R1  k12 
=λ  kF 2 − kF1 
kc1k11  k22 
On the other hand, when loop 2 is opened, the area under e1becomes:
∞
τ R' 1 k F 1
∫ e dt = − k
o
1 '
0 c1 k11

Thus,

8
 ∞

∫ e dt 1
τ R1 kc' 1  k F 2 k12  τ R1 kc' 1
0
=− ×  − 1  λ = × β1 = f1 β1
∞
τ R' 1 kc1  k F 1k22  τ R' 1 kc1
∫ e dt
o
1
0

Similarly, we have:
∞

∫ e dt 2
τ R 2 kc' 2
0
= × β2 = f2 β2
∞
τ R' 2 kc 2
∫ e dt
o
2
0

Notice that the PI parameters in the interacting loops are used to be more conservative
than those in single loops. In another words,
f1 ≥ 1; f2 ≥ 1
The multi-loop control should be beneficial when the sum of absolute values of the

Remarks:
1. If λ is assumed not vary with frequency, and the process under study is FOPDT,
λ>1, f1 lies in the range 1< f1 <2, while 0.5<l<1, f1 lies in the range 1< f1 <3.
2. When f1 =1, β is equal to the ratio of response areas.
3. If β is small and f1 =is close to one, then the interacting control is favored for that
particular disturbance.
4. If β is large, the interacting control is un favorable for that particular disturbance.

The Relative Gain for Non-square Multivariable Systems

(J.C. Chang and C.C. Yu, CES Vol.45, pp. 1309-1323 1990)

Consider a non-square MV system.

ym×1 ( s ) = Gm×n ( s )un×1 ( s )
Define Moore-Penrose pseudo-inverse of the matrix G ( s) as:

G + ( s) = ( GT G ) GT ( s)
−1

Then, under close-loop control, the steady-state control input will be:
 ∂ui  +
u = G + (0) y d and   = gij (0) .
∂y
 j CL
Thus, the non-square relative gain is defined similarly to the square RGA, that is:

9
 −1
 ∂y   ∂yi   + T
Λ% =  i     = G (0) ⊗ G (0) 
 ∂u
 j OL  ∂u j CL 

Properties of the non-square RGA

1. Row sum of Λ% :
T
 n n n

RS = [ rs (1), rs (2), L , rs (m) ] =  ∑ λ%1 j , ∑ λ% 2j , L , ∑ λ% mj , ;
 j =1 j =1 j =1 

Where, rs(i) = G (0)G + (0)  ii

T
 n n n

CS = [ cs (1), cs (2), L , cs (n)] =  ∑ λ% j1 , ∑ λ% j 2 , L , ∑ λ% jn , = [1, 1, L, 1]
T
2.  j =1 j =1 j =1 

Where, cs (i ) = G + G (0)  ii ; (Note: G + G = (GT G ) −1 GT G = I )

3. 0 ≤ rs(i) ≤ 1, ∀i = 1, 2,L, m
m n
4. ∑ rs(i) = ∑ cs(i) = n
i =1 j =1

m m n n m n
Note: ∑ rs(i) = ∑∑ λ% = ∑∑ λ% = ∑ cs ( j ) = n
i =1 i =1 j =1
ij
j =1 i =1
ij
j =1

5. Non-square RGA is invariant under input scaling, but is variant under output
scaling:

( GS ) ⊗ (GS )+  = (G ⊗ G+ ) ( SG ) ⊗ ( SG )+  ≠ (G ⊗ G+ )
T T T T

6. Let P1 and P2 are permutation matrices. Then, Λ% ( PGP

1
%
2 ) = P1 Λ (G ) P2

A. Multi-loop BLT-Tuning:

I. BLT-1 method:

a. Calculate the Ziegler-Nichol settings for each PI controller by using the

diagonal element of G, i.e. gi,i.
b. Assume a detuning factor “F”, and calculate controller settings for loops.

kc ,i = k ZN ,i / F ; τ R ,i = (τ R ,i )ZN F

c. Define: W(iω ) = −1 + det  I + G( iω )Gc ( iω ) 

10
 d. Calculate the closed-loop function Lc(iω):

W( iω )
Lc (iω ) = 20 log
1 + W( iω )

e. Calculate the detuning factor F until the peak in the Lc log modulus curve is
equal to 2N, that is:

 W( iω ) 
Lcm = Max 20 log  = 2N
ω
 1 + W( iω ) 

II. BLT-2

a. Find BLT-1 PI controllers.

b. Choose a second detuning factor FD. FD should be greater than one.
c. Compute τD,j as:

τ D, j =
(τ )
D , j ZN

FD

d. Calculate W(iω) and Lc(iω).

e. Change FD until Lmax

C is minimized, maintaining FD > 1. The trivial case may

C is minimized for FD = ∞ , i.e., no derivative action.

result where Lmax

f. Reduce F in the P and I modes, until Lmax

C = 2N .

III. BLT-3

The objective is to estimate the level of imbalance in detuning the BLT-1

controller and compensate for it.
Consider the PI controller:

 1
t

u j = u j (0) + kC , j  e j +
 ∫
τ R, j 0
e j dt  ;

u j (0) = 0
 
At steady state,

11
 ∞
kC , j
lim u j (t )  =
t →∞ τ R, j ∫ e (t )dt
0
j

So,
∞
τ R , j u j (∞)
∫ e (t )dt =
0
j
kC , j

Notice that:
u (∞) = G −1 (0) R − G −1GL (0) d (∞)

For unit step set-point input:

u j (∞) = G −1 (0)[0,..., 0,1, 0,..., 0]T

= [ g i , j (0); i, j = 1,..., N ][0,..., 0,1, 0,..., 0]T

For unit step load disturbance:

N
ui (∞) = i th row of G −1 (0)GL (0) = ∑ [ gi , j (0) g L, j (0)]
j =1

Then, ITEj becomes:

u j (∞)τ R , j
ITE j =
kC , j

Let,
N ITE j
Sj = ∑ + ITE j −load
i =1 N

τ R, j  N g j ,i (0) 
Sj = × ∑ + g j ,i (0) g L,i (0) 
kC , j  i =1 N 

Let S max = Max S j

j

Smax
Fj = F
Sj

The PI controller parameters becpme:

kc ,i = kZN ,i / Fj ; τ R ,i = (τ R ,i ) ZN Fj

IV. BLT-4

12
 a. BLT-3 is used to get individual PI controllers as described above.
b. BLT-2 procedure is used with individual FD factors for each loop:

S max
FD , j = FD
Sj

V. Tyreus Load-Rejection Criterion (TLC)

The best variable pairing is the one that gives the smallest magnitudes for each
element of X,(i.e. Xi) of the following:

X ( iω ) = ([ I + GGC ]−1 GL L )
( iω )

VI. Summary

BLT-1--- PI, BLT-2---PID,

equal Fi equal Fi

BLT-3---PI, BLT-4-----PID,
unequal Fi unequal Fi

B. Parallel-design method---Modified Z-N methods for

TITO Processes

This method is based on A modified Z-N method for SISO control system. To derive
this modified Z-N method, ageneral formulation is to start with a given point of the
Nyquist curve of the process:
j ( −π +ϕ p )
G p ( jω ) = rp e (1)

And to find a regulator GR

13
   1 
GR ( jω ) = k  1 + j  ωτ D −  (2)
  jτ Rω  

To move this point to B = rs e j ( −π +ϕs ) (3)

An amplitude margin (i.e. gain margin) design corresponding to ϕ s = 0 and

1
rs = .
Am
A phase margin design corresponds to rs = 1 and ϕ s = ϕm
j ( −π +ϕ p +ϕ R )
From Eqs.(1)~Equ.(3), we have: rs e j ( −π +ϕs ) = rp rR e , so that

rs
rR = and ϕ R = ϕ s − ϕ p
rp
In other words,

  1  j (ϕ R )
GR ( jω ) = k  1 + j  ωτ D −   = rR e = rR cos ϕ R + jrR sin ϕ R
  jτ Rω  

Or,

 1 
k = rR cos ϕ R =
rs
rp
(
cos ϕ s − ϕ p ) and  ωτ D − (
 = tan ϕ s − ϕ p
τ Rω 
)

The gain is uniquely determined. Only one equation determines τ R and τ D .

Let τ D = ατ R , where α is often chosen as α ≈ 0.25 . Another method to specify α

is as follows:

0.413 g (0)
α= , where κ =
3.302κ + 1 g ( jωc )

 1 
From  ωτ D −
τ ω
−1
( )
 = tan ϕ s − ϕ p , τ D can be solved to obtain:
 R 

1 
τD = − tan(ϕ s − ϕ p ) + 4α + tan 2 (ϕ s − ϕ p )  and
2ω  
1
τR = τD
α

Consider a stable 2 × 2 process :

14
 y1 ( s)   g11 ( s) g12 ( s )   u1 ( s ) 
 y ( s )  =  g ( s ) g ( s )  u ( s ) 
 2   21 22  2 

 c1 ( s )  c1 ( s ) 0 
c ( s )  =  0 c2 ( s) 
 2  

c2 g12 g 21 g g
g1 = g11 − = g11 − −112 21
1 + c2 g 22 c2 + g 22
g12 g 21
g 2 = g 22 −
c1−1 + g11

Let

Ai = rai e j ( −π +ϕai ) = gi ( jωi )

Bi = rbi e j ( −π +ϕbi ) = gi ( jωi )ci ( jωi )

  1 
ci ( jω ) = k 1 + j  ωτ Di +   ; i = 1, 2
 jτ Riω  
 

Take PI controller as example.

ci ( jω ) = kci (1 − j tan(ϕbi − ϕ ai ) ) ; i = 1, 2

And, gi ( jωi )kci = cos(ϕbi − ϕai )rbi e j ( −π +ϕai )

15
 rai e j ( −π +ϕia ) ⋅ kci (1 − j tan(ϕbi − ϕai )) = rbi e j ( −π +ϕia )
⇓
rbi j (ϕai −ϕbi ) rbi r
e = cos(ϕai − ϕbi ) + j bi sin(ϕai − ϕbi ) = kci (1 − j tan(ϕbi − ϕ ai ))
rai rai rai
⇓
rbi
kci = cos(ϕai − ϕbi )
rai
rbi
kci ⋅ gi ( jω ) = cos(ϕ ai − ϕbi ) ⋅ rai e j ( −π +ϕia ) = rbi cos(ϕ ai − ϕbi ) ⋅ e j ( −π +ϕia )
rai

By setting i equal one and two, one will obtain two equations with kc1 and kc2 as
unknowns, and, thus, can be solved. But, there are very tedious procedures to find the
controller gains (such as:such kc1 and kc2) and frequency ω11 and ω22 that satisfy the
phase criteria. (see the reference: I&EC Res. 1998, 37, 4725-4733, Q-G Wang, T-H
Lee, and Y. Zhang)

C. Independent design method

---IMC Multi-loop PID Controller

16
 f1 [(g11) -]-1
_
G
_ f2 [(g22)-]-1

g11 _

g22 _

GC ,i = ( Gi ,i )− f i ; i = 1,..., n
−1

The stability is guaranteed for any stable IMC filter that satisfies either of the
following:

g i ,i (iω )
f i (iω ) < f R*,i (iω ) = ; i = 1, 2,..., n
∑
j , j ≠i
gi , j (iω )

g i ,i (iω )
f i (iω ) < f C*,i (iω ) = ; i = 1, 2,..., n
∑
j , j ≠i
g j ,i (iω )

Imc Row interaction measure [Economou and Morari]

1
∑
j , j ≠i
g i , j (iω )
Ri (iω ) = = ; 0 ≤ω ≤ ∞
1 + f R*,i (iω ) ∑g j
i, j (iω )

1
∑
j , j ≠i
g j ,i (iω )
Ci (iω ) = = ; 0 ≤ω ≤ ∞
1 + fC*,i (iω ) ∑g
j
j ,i (iω )

For significant interaction: 0.5 ≤ Ri , Ci ≤ 1 ⇒ f * <1

17
 For small interaction: 0.0 ≤ Ri , Ci ≤ 0.5 ⇒ f * >1

D. Chien-Huang-Yang’s multi-loop PID---with no

proportional and derivative kicks

1. Controllers for SISO loop:

 1 
Controller: u ( s ) = kC  − y ( s ) + [r ( s ) − y ( s)] − τ D sy ( s) 
 τ Rs 

y k /(τ s )G p
= C R
r 1 + kC /(τ R s )G p

a. Time constant dominant processes:

Re − Ls
GP = ; R = slope of the initial unit step response
s
Re− Ls R(1 − Ls)
GP = ≈
s s
y 1 − Ls 1 − Ls
= ≈ 2 2
r  τR  τ C s + 1.414τ C s + 1
 − τ R L  s 2 + (τ R − L ) s + 1
 RkC 
(1.414τ C + L)
⇒ kC = ; τ R = 1.414τ C + L
R(τ + 1.414τ C L + L2 )
2
C

b. Deadtime dominant processes:

k P e− Ls k P (1 − Ls )
GP = ≈
τ s +1 τ s +1

18
 y 1 − Ls
=
r  τ Rτ   τ 
 − τ R L  s2 +  R + τ R − L  s + 1
 kC k P   kC k P 
1 − Ls
≈
τ C s + 1.414τ C s + 1
2 2

1 −τ C2 + 1.414τ Cτ + Lτ
⇒ kC = ;
k P τ C2 + 1.414τ Cτ + L2
−τ C2 + 1.414τ Cτ + Lτ
⇒ τR =
τ +L

Derivation of the PID controller parameters is similar to the above PI

derivations except that the deadtime approximation:

1 − 0.5 Ls
e − Ls ≈
1 + 0.5 Ls

19
20
21
 2. Controllers for multi-loop system

 y  k k  g1,1
At ω → 0 ;   = g1,1 1 − 1,2 2,1  =
 u1 loop 2 closed  k1,1k2,2  RGA(λ )

 y
At ω → ∞ ;   = g1,1
 u1 loop 2 closed

a. For RGA>1, multi-loop controller tuning based on the process

model in the main loop should provide satisfactory closed loop
results. This is because:
b. For RGA < 1,

kC ,i = ( kC )based on main loop RGA(λi ,i )

(τ )
R ,i based on main loop
τ R ,i =
RGA(λi ,i )

τ D ,i = (τ D ,i )based on main loop RGA(λi ,i )

The closed-loop time constant is chosen according to the value of L/τ

in three different ranges, that is: L/τ < 0.2, 0.2 < L/τ < 0.5, and L/τ >
0.5.
For details, see the original paper.
.
IX. Robustness of Closed-loop System.

The final pairing and the controller tuning is checked for robustness by
plotting DSO and DSI as functions of frequency, [Doyle and Stein]. The
singular values below 0.3-0.2 indicate a lack of stability robustness.

DSO(iω ) = σ [ I + ( GGC ) ]( iω )
−1

DSI ( iω ) = σ [ I + ( GC G ) ]( iω )
−1

22
E. Design Method based on Passivity

1. Hardware simplicity and relative effortlessness to achieve failure tolerant design,

multi-loop control is the most widely used strategy in the industrial process
control.
2. Current multi-loop control design approaches can be classified into three
categories: detuning methods (Luyben, 1986), independent design methods
(Skogestard and Morari, 1989), and sequential design methods (Mayne, Chiu and
Arkun, 1992).
3. Loop interactions have to be taken into considerations, as they may have
deteriorating effects on both control performance and closed-loop stability.
4. It is desirable if the multi-lop control system is decentralized unconditionally
stable (i.e., any subset of the control loops can be independently to an arbitrary
degree or even turned off without endangering close-loop stability.
5. Independent design is based on the basis of the paired transfer function while
satisfying some stability constraints due to process interactions.
6. Perhaps the mostwidely used decentralized stability conditions are those
µ-interaction measure.
7. Passivity Concept:
The rate of change of the stored energy in the tank is less than the power supplied
to it.
Inlet Flowrate
Fi

Fo
h Outlet Flowrate

Potential energy stored in the tank: S (h) = 12 Ahρ gh = 12 Aρ gh 2

Increment of potential energy per unit time: w(t ) = ρ Fi (t ) gh(t )

The rate of change of the storage function:
dS
= −Cv ρ gh h + ρ gFi h = −Cv ρ gh h + w < w ∀h > 0
dt
The rate of change of the stored energy in the tank is less than the power
supplied to it. Therefore this process is said to be strictly passive.

23
Passive(Willems 1972): if a non-negative storage function S(x) can be found s.t.:
t
S(0)=0 and S ( x) − S ( x 0 ) ≤ ∫ yT (τ )u (τ )dτ for all t>t0≥0, x0, x∈ X, u∈ U.
t0

t
Strictly passive: if S ( x) − S ( x 0 ) < ∫ yT (τ )u (τ )dτ
t0

Where, y is the output of a system, u is the input to the system.

 KYP Lemma
 Nonlinear control affine systems (Hill & Moylan 1976)
x& = f ( x) + g ( x)u
y = h( x )
where x ∈ X ⊂ R n , u ∈ U ⊂ R m , y ∈ Y ⊂ R m
The process is passive if
∂S T ( x )
Lf S ( x) = f ( x ) ≤ 0,
∂x
∂S T ( x )
Lg S ( x ) = g ( x ) = hT ( x )
∂x
 KYP Lemma
A linear system (Willems 1972) G(s):=(A,B,C,D) is passive if there exists a
positive definite matrix P such that:

 AT P + PA PB − C T 
 T ≤0
 B P − C − D − D 
T

The system is strictly passive if

 AT P + PA PB − C T 
 T <0
 B P − C − D − D 
T

Definition:
An LTI system S: G(s) is passive if :
(1) G(s) is analytic in Re(s)>0;
(2) G(jw)+G*(jw)≥0 for all that jw is not a pole of G(s);
(3) If there are poles of G(s) on the imaginary axis, they are non-repeated and the
residue matrices at the poles are Hermitian and positive semi-definite.
G(s) is strictly passive if:
(1) G(s) is analytic in Re(s) ≥ 0;
(2) G(jw)+G*(jw)>0 ∀ω ∈ ( −∞, ∞) .

24
Theorem 1: For a given stable non-passive process with a transfer function matrix
G(s), there exists a diagonal, stable, and passive transfer function matrix
W(s)=w(s)I such that H(s)=G(s)+W(s) is passive.
[Proof]:

λmin ( H ( jω ) + H * ( jω )) = λmin (G ( jω ) + G* ( jω ) + (W ( jω ) + W * ( jω ))

Since both (G+G*) and (W+W*) are Hermitian, from the Weyl inequality, we
have:

λmin ( H ( jω ) + H * ( jω )) ≥ λmin (G ( jω ) + G* ( jω )) + λmin (W ( jω ) + W * ( jω ))

=λmin (G ( jω ) + G* ( jω )) + 2 Re(W ( jω ))

Thus, if:
1
Re(W ( jω )) ≥ λmin (G ( jω ) + G* ( jω ))
2
H(s) can be render passive. On the other hand, if
1
Re(W ( jω )) > λmin (G ( jω ) + G* ( jω ))
2
H(s) will be strictly passive.

Properties of Passive Systems:

 A passive system is minimum phase. The phase of a linear process is within
[-90º, 90º]
 Passive systems are Lyapunov stable
 A passive system is of relative degree < 2
 Passive systems can have infinite gain (e.g., 1/s)

Passivity Theorem :

If G1 is strictly passive and G2 is passive, then the closed-loop system is L2

stable.

 A strictly passive process can be stabilized by any passive controller

25
 (including multi-loop PID controllers) even if it is highly nonlinear and/or
highly coupled
 Control design based on passivity
 Excess or shortage of passivity of a process can be used to analyse whether
this process can be easily controlled
 Passivity based controllability study

A non-passive process can be made passive using feedforward and/or feedback

passification:

Gff Gfb

G G

The excess or shortage of passivity can be quantified using:

 Input Feedforward Passivity (IFP) (Sepulchre et al 1997) - If a system
G with a negative feedforward of νI is passive, then G has excessive
input feedforward passivity, i.e., G is IFP(ν).

 Output Feedback Passivity (OFP) (Sepulchre et al 1997) - If a system

G with a positive feedback of ρI is passive, then G has excessive
output feedback passivity, i.e., G is OFP (ρ).

Agin, use the following figure:

If G1 is IFP(ν) and G2 is OFP(ρ), then the closed-loop system is stable if ρ+ν>0.

In other words, a processs shortage of passivity can be compensated by another
process’s excess of passivity.

 Passivity Index
The excessive IFP of a system G(s) can be quantified by a frequency dependent

26
passivity index
∆
1 
ν F [G ( s ), ω ]=λmin  [G ( jω ) + G *( jω )] 
2 
Assume the true process is GT ( s ) = G ( s ) + ∆( s)
The passivity index of the true process can be estimated as
1 1 
ν (GT ( jω )) = −λmin   ∆( jω ) + ∆* ( jω )  + G ( jω ) + G* ( jω )  
2 2 
1  1 
≤ −λmin   ∆( jω ) + ∆* ( jω )   − −λmin  G ( jω ) + G* ( jω )  
2  2 
=ν (G ( jω )) +ν (∆ ( jω ))

Properties of the Passivity Index

1. Comprises gain & phase information of the uncertainty

j
Maximum gain

Passivity index

σ
∆(σ)
 

2. Always no greater than the maximum singular value.

ν F [∆( s), ω ] ≤ σ max [∆( jω )] for any ω ∈ R

27
 Passivity Theorem 2: If the multivariable process is strictly passive, then the
closed-loop system is stable if the multi-loop controller is passive.

Theorem 1: A closed-loop system comprising a stable subsystem G(s) and a

decentralized controller K(s)=diag(ki(s)), w(s) is a stable and minmum phase, and

ν (W ( jω )) < −ν (G + ( jω ))

Then the closed-loop system will be decentralized unconditional stable, if

K(s)=diag{ ki' ( s ) } is passive, where,

ki' ( s ) = ki+ [1 − w( s)ki+ ( s)]−1 and ki+ = U ii ki

−1
K ' ( s ) = U −1K ( s)  I − w( s )U −1K ( s) 

Notice that the above figure is equivalent to the one in the following:

Rescaling of the system

Let D(s) be a diagonal, real and constant scaling matrix.

The scaling matrix D(s) is to make

ν ( D −1G + D ( jω )) < ν (G + ( jω ))

and

28
 +
D −1G + (0) D + D −1 G + (0)  D > 0

Design procedures:

1. Find matrix U and calculate G + ( s) .

2. Check the pairing. Examine the proposed pairing using DIC condition:
T
G + (0) M + M G + (0)  > 0

3. Use matrix M obtained in the step 2 to derive D, D = M 1/ 2

4. Calculate ν ( D −1G + ( jω ) D ) for different frequency points. These frequency

points form a set Ω ∈ [ 0, ωE ] where ωE is the frequency which is high enough

sych tant ν ( D −1G + ( jω ) D) → 0 for ω > ωE .

5. For each loop of the controller, solve problem:

min(−γ i )
kci ,τ Ri

such that

1 γi
<1
 1  jω
1 + Gii+ ( jω )kc+,i 1 + 
 jτ R,iω 

and

kc+,iν s (ω )
τ R2 ,i ≥ , ∀ω ∈ R, i = 1,L , n
1 − kc+,iν s (ω )  ω 2
 

6. Obtain the final controller settings: kc ,i = U ii kc+,i

This method is limited to open-loop stable processes.

29
 ν F (∆( s ), ω ) ≥ −ν F (W ( s ), ω ), ∀ω ∈ R

Robust Stability Condition

If the uncertainty is passive, then the controller is only required to render system T
strictly passive to achieve robust stability even if ∆ is very large.

If the uncertainty’s passivity index is bounded by

ν F ( ∆( s ), ω ) ≥ −ν F (W ( s), ω ) , ∀ω ∈ R

where W(s) is minimum phase, the closed-loop system will be robust stable if
system

T ( s )[ I − W ( s )T ( s)]−1

is strictly passive.

The basic idea:

1. Characterise the uncertainty in terms of passivity using IFP or OFP.
2. Derive the robust stability condition for systems with uncertainties bounded by
their passivity indices.
3. Develop a systematic procedure to design the robust controller which satisfies
the above stability condition.

Passivity Based Robust Control Design

 Blended approach
 Design a controller that satisfies the small gain condition at high
frequencies and satisfies the passivity condition at low frequencies
(Bao, Lee et al 1998)
 Based on the bilinear transformation

30
  Multi-objective control design
 Design a controller that satisfies the passivity condition for robust

stability and achieves H∞ control performance (Bao, Lee et al 2000,

2003)
 Based on KYP lemma and Semi-Definite Programming

Example:

 0.126e −6 s − 0.101e −12 s 

 
G ( s) =  60s +−18 s (48s + 1)(45s + 1) 
−8 s
 0.094e − 0.12e 
 38s + 1 35s + 1 

Passivity index

0.0.0404

P
0.0.0202
as
0.0
si 0.0 0

vit
-0.0
-0.-0.020

-0-0.03.0
110-40
- 10-2
1 1 100 10+2
1 10+41 4
(rad/min

F. Design by Sequential Loop Closing

Advantages of sequential design:

1. Each step in the design procedure involves designing only one SISO controller.
2. Limited degree of failure tolerance is guaranteed: If stability has been achieved

31
 after the design of each loop, the system will remain stable if loop fail or are taken
out of service in the reverse order of they were designed.
3. During startup, the system will be stable if the loops are brought into service in the
same order as they have been designed.
4.
Problems with sequential design:
1. The final controller design, and thus the control quality achieved, may depend on
the order in which the controllers in the individual loops are designed.
2. Only one output is usually considered at a time, and the closing of subsequent
loops may alter the response of previously designed loops, and thus make iteration
necessary.
3. The transfer function between input uk and output yk may contain RHP zeros that
do not corresponding to the RHP zeros of G(s).

Notations:

1. G(s): the n × n matrix of the plant, G ( s ) = {gij ( s ); i, j = 1,L , n}

2. C ( s) = diag{ci ( s); i = 1,L , n}

3. S = ( I + GC )−1 ; H = I − S = GC ( I + GC ) −1

4. G% = diag{gii ( s ); i = 1,L , n}

1
5. S% = diag{si ( s ); i = 1,L , n} = diag{ ; i = 1,L , n}
1 + gii ci
g c
6. H% = diag{hi ( s ); i = 1,L , n} = diag{ ii i ; i = 1,L , n}
1 + gii ci

% −1 = {γ ; i, j = 1,L , n}
7. Γ = GG ij

% −1G
8. CLDG = GG d

9. E = (G − G% )G% −1

G M C M
10. G =  k  ; C= k ;
 L O  L O

11. Sk = ( I + Gk Ck ) ; H k = Gk Ck ( I + Gk Ck )
−1 −1

Hk 0  Sk 0
12. Hˆ k =   ; Sˆk =  ; i = k + 1, K + 2, L N
 0
%
hi  0 s%i 

32
 S = ( I + GC ) −1 = [ I + GC
% + (G − G% )C ]−1

{ )}
−1
( ) (
−1
=  I + (G − G% )C I + GC
%  I + GC
%
 

{ )}
−1
( )  (
−1 
=  I + (G − G% )G% −1GC
% I + GC
% %
I + GC


( ) ( I + EH% ) ( )
−1 −1 −1
%
= I + GC = S% I + EH%

Design procedures:
) ) )
( )
−1 −1
In each of the following step, S = S k ( I + Ek H k ) ; Ek = (G − Gˆ k ) Gk

Determine ci such that W p SWD is minimized.

Step 0. Initialization. Determine the order of loop closing by estimating the

required bandwidth in each loop. Also estimate the individual loop designs
in terms of H% .
Step 1. Design of controller c1 by considering output 1 only. In this case, we have

Gˆ k = G% k and Hˆ k = H%

Step k. Design of controller ck by consider outputs 1 to k. Here,

Gˆ k = diag{G% k , gii }; i = k + 1, k + 2,L , n and

Hˆ k = daig{H k −1, h%i } ; i = k , k + 1,L , n

33
Sequential Design Using Relay feedback Tests of Shen and Yu

The relay feedback system for SISO auto-tuning is as shown in the follwing figure:

When constant cycles appear after the system has been activated, the ultimate gain
and ultimate frequency of the open-loop system can be approximated by measuring
the magnitude and period (see the following figure) and by the following equations:

4h 2π
Ku = ; ωu =
πa Pu
The Z-N tuning method can be used to determine the controller parameters:

PI Controller: K c = 0.45K u , τ R = Pu /1.2,

PID Controller: K c = 0.60 Ku , τ R = Pu /1.2, τ D = 1.25 Pu
Or, use the Tyreus-Luyben’s formula to give more conservative response:
PI Controller: K c = K u / 3.2, τ R = 2.2 Pu ,
PID Controller: K c = K u / 2.2, τ R = 2.2 Pu , τ D = Pu / 6.3
To avoid the difficult mathematics envolved in the formulation of sequential
design, Shen and Yu suggested to use the relay-feedback test as shown in the
following figure:

34
 The controller for a 2 × 2 system is suggested:

PI Controller: K c = K c, ZN / 3, τ R = 2 Pu

Analysis:
The sequential design is derived by considering the multi-loop control system as
coupled SISO loops. For a 2 × 2 system as example, the equivalent SISO loops are:
1
g1 ( s) = g1,1 ( s ) {1 − (1 − ) h ( s)}
λ (s) 2
1
g 2 ( s) = g 2,2 ( s ) {1 − (1 − ) h ( s)}
λ (s) 1

g C,i gi ,i
Where, hi (s) = ; i = 1, 2
1 + gC ,i gi ,i

Notice that, if there is damping in g1 or g 2 , this damping should come from either
h1 or h2 . According to tis study, a closed system having an FOPDT process and a
modified ZN tuned PI controller will result in a closed-loop system (i.e. h1 and h2 )
having damping factor greater than 0.6. It is thus postulate that the open-loop transfer
functions g1 ( s ) and g 2 ( s ) can be approximated by:

kp τ p 2 s + 1 −θ s
G ( s) = ⋅ ⋅e
τ 2 s 2 + 2τζ s + 1 τ p1s + 1

Then, the stability region of the equivalent SISO loops are explored with the

parameters: τ p1 ,τ p = 0 ~ 10, k p = 1, τ = 5, ζ = 0.1 ~ 1, θ / τ = 0.02 ~ 0.2 . The results

35
are given in the following figure. It can be seen that the modified ZN tuning formula
proposed greatly improve the stability.

On the other hand, the convergence of the sequential design for the multi-loop
controller is formulated as the problem of finding the roots of simultaneous algebraic
equation using sequential iterations.
The simultaneous equations are obtained from the conditions of phase crossover for
the two loops, that is:
Im  g1 ( jωu ,1 , jωu ,2 ) 
F1 ( jωu ,1 , jωu ,2 ) = tan −1 = −π
Re  g1 ( jωu ,1 , jωu ,2 ) 
Im  g 2 ( jωu ,1 , jωu ,2 ) 
F2 ( jωu ,1 , jωu ,2 ) = tan −1 = −π
Re  g 2 ( jωu ,1 , jωu ,2 ) 

The convergence of the iteration is guaranteed by a sufficient condition of the

following:
 ∂F1   ∂F2 
   
 ∂ωu ,2 ωu ,1  ∂ωu ,1 ωu ,2
<1
 ∂F1   ∂F2 
   
 ∂ωu ,1 ωu ,2  ∂ωu ,2 ωu ,1

36
 The procedures of this proposed sequential design are summarized with the flow
chart as shown.

37
Design of Multi-loop control systems

Consider a single loop system as shown in Fig.1.

Fig.1

Suppose controller is fixed, substantial changes in K p invariably lead to deteriorate the

control system response (see Figure 2).

Fig.2

Now consider a 2 × 2 control system in Fig. 3

Fig.3

Let us consider open loop 1 (Fig.4)

1
 Fig.4

One may notice that there are two transmission paths from m1 to x1 .let us define a
relative gain from m1 to x1 .as:
Gain m1 − x1 , loop 2 open
λ=
Gain m1 − x1 , loop 2 close
When λ = 0.5 , the responses of x1 to a unit step input at m1 is shown in Fig. 5

Fig.5

In this case, when loop 2 is closed, the open loop gain of m1 − x1 becomes doubled.
The increase in the loop gain results in more oscillation in the closed loop response as
shown.
On the other hand, when λ = 2 , the open loop and closed loop responses are also

2
given in Fig. 6.

In this latter case, the open-loop gain decrease when loop 2 is switched from open to
close. As a result, the close of loop 1 leads the system to a more sluggish response to
the r1 input.
The increase or decrease of the loop gain is a result of closing loop2, and , hence, is
considered loop interaction. From the above example, λ is a measure of such
interaction and is named as relative gain of loop 1. You may also find the other
relative gain for loop2. But, in this case, the two relative gains will be equal.

Algebraic Properties of the RGA

1. ∑g
j
ij g ji = ∑ λ
j
ij = 1, ∀i

2. P1Λ{G} P2 = Λ{P1G P2}, P1 and P2 are two permutation matrices.

3. Λ{G} = Λ{S1G S2}, S1 and S2 are two diagonal matrices.
4. If transfer matrix, G, is diagonal or triangular, then: Λ{G}=I.
[Proof]:
 x 0 0...0 
 x x 0...0 
Let, G =  
.... 
 
 x x x... x 
Then,

3
  x 0 0...0 
 x x 0...0 
G −1 =G = 
.... 
 
 x x x... x 
Thus, g ij g ji = 0 = λ ij, ∀j ≠ i

and, g ii g ii = 1 = λ ii , ∀j = i

So, Λ =I

dg ji dg ij
5. = −λ ij
g ji g ij

adj[ A] ( −1) i + j det[G ij ]

g ij = =
det[G ] det[G ]
d det[G ]
( −1) i + j det[G ij ]
dg ji dg ij det[G ij ]2
=− 2
= 2
= − g 2ji
dg ij det[G ( g ij )] det[G ( g ij )]

dg ji dg ij dg ij
= − g ji dg ij = − g ji g ij = −λ ij
g ji g ij g ij

dλ dg ij dλ λ ij − 1 dg ij
= (1 − λ ij ) =
ij ij
5. , and
λ ij g ij λ ij λ ij g ij

[Proof];

λ ij = g ij g ji ⇒ dλ ij = dg ij g ji + g ij dg ji

dλ dg ij g ji + g ij dg ji dg ij dg ji dg ij
⇒ = = +( ) = (1 − λ ij )
ij

λ ij g ij g ji g ij g ji g ij

or,

dλ  dg ij  dg ji 1 dg ij  λ ij − 1 dg ij
⇒ = + = (1 − =
ij
 g  g
) 
λ ij  ij  ji λ ij g ij  λ ij  g ij

4
 RGA-implications:

1. Pairing loops on λij values that are positive and close to 1.

2. Reasonable Pairings: 0.5 < λij< 4.0

3. Pairing on negative λij values results in at least one of the following;

a. Closed loop system is unstable,
b. Loop with negative λij is unstable,
c. Closed loop system becomes unstable if loop with negative is λij turned off.

4. Plants with large RGA-elements are always ill-conditioned. (i.e., a plant with a
large γ(G) may have small RGA-elements)

5. Plants with large RGA-elements around the crossover frequency are fundamentally
difficult to control because of sensitivity to input uncertainties.
-----
decouplers or other inverse-based controllers should not be used for plants
with large RGA-elements.

6. Large RGA-element implies sensitivity to element-by-element uncertainty.

7. If the sign of RGA-element changes from s=0 to s= ∞ , then there is a RHP-zero

in G or in some subsystem of G.

8. The RGA-number can be used to measure diagonal dominance:

RGA-number = || Λ(G)-I ||min.
For decentralized control,, pairings with RGA-number at crossover frequency
close to one is preferred.

9. For integrity of whole plant, we should avoid input-output pairing on negative

RGA-element.

10. For stability, pairing on an RGA-number close to zero at crossover frequency is

preferred.

5
 The Relative Disturbance Gain (RDG)

Ref: Galen Stanley, Maria Marino-Galarraga, and T. J. McAvoy, Shortcut

Operability Analysis. 1. The relative disturbance gain, I&EC, Process Des. Dev.
1985,24, 1181-1188

The use of RDG:

1. To decide if interaction resulting from a disturbance is favorable or unfavorable.
2. To decide whether or not decoupling should be used and what type of decoupling
structure is best.

y1 = k11m1 + k12 m2 + k F 1d
y2 = k21m1 + k22 m2 + k F 2 d

 ∂m1  k
 ∂d  = − F1
  y1 , m2 k11

 ∂m1  is derived when both y1 and y2 are held still:

 ∂d 
  y1 , y2

y1 = k11m1 + k12 m2 + k F 1d = 0 (2)

y2 = k21m1 + k22 m2 + k F 2 d = 0
so that:

1
m2 = [ −k21m1 − k F 2 d ] (3)
k22
Substitute Eq.(3) into E.(2), we have:

 k12 k21   k12 k F 2 

k
 11 − m +
 1  F1k − d = 0
 k 22   k 22 

Thus,
k12 k F 2
−kF1 + (4)
 ∂m1  k22 k k − k F 1k22
 ∂d  = = 12 F 2
  y1 , y2 k k k11k22 − k12 k21
k11 − 12 21
k22

6
So,

 ∂m1 
 ∂d 
  y1 , y2 k k k − k F 1k22 k11 k22 k12 k F 2 − k F 1k22 (5)
β1 = = − 11 × 12 F 2 = ×
 ∂m1  k F 1 k11k22 − k12 k21 k F 1k22 k11k22 − k12 k21
 ∂d 
  y1 , m2
k12 k F 2 − k F 1k22 k11k22  k k 
=- × = 1 − 12 F 2  λ
k F 1k22 k11k22 − k12 k21  k F 1k22 

Similarly, we have:

 ∂m1 
 ∂d 
  y1 , y2  k21k F 1 
β1 = = 1 − λ
 ∂m1   k k
F 2 11 
 ∂d 
  y1 , m2

k F 2 k12 β λ − β1 k F 2 λ − β1 k22
= 1− 1 = ⇒ = ×
k F 1k22 λ λ kF1 λ k12

k F 1 λ − β 2 k11
Similarly, ⇒ = ×
kF 2 λ k21

So, kF 2 kF 1  λ − β 2 k11   λ − β1 k22 

× =1=  × × × 
kF1 kF 2  λ k21   λ k12 

 λ − β2   λ − β1  k11 k22 1
⇒  λ   λ  = k k = 1− λ
   21 12

or,

( β1 − λ )( β 2 − λ ) = λ (λ − 1)
λ (λ − 1) λ (λ − 1) ( β − 1) λ 1 − β1
( β2 − λ ) = ⇒ β2 = +λ = 1 = λ
( β1 − λ ) ( β1 − λ ) ( β1 − λ ) λ − β1
It can be shown that:

7
 Multi-loop e1 area
∝ β1 and
SISO idealy decoupled e1 area
Multi-loop e 2 area
∝ β2
SISO idealy decoupled e 2 area

e1 F2 G12 Gc 2 − F1 (1 + G22 Gc 2 ) (6)

=
d (1 + G11Gc1 )(1 + G22 Gc 2 ) − Gc 2 Gc1G12 G21

If d is a unit step, then the area under e1 curve is given as:

k k k k
∞ k F 2 k12 c 2 − F 1 22 c 2
τ R2 τ R2 k F 2 k12 − k F 1k22
∫0 e1dt = lim
s →0
e1 ( s) =
kc1k11 kc 2 k22 kc 2 kc1
=
kc1k11k22  k12 k21 
− k k 1−
τ R1 τ R 2 τ R 2 τ R1 12 21 τ R1  k11k22 
τ R1  k12 
=λ  kF 2 − kF1 
kc1k11  k22 
On the other hand, when loop 2 is opened, the area under e1becomes:
∞
τ R' 1 k F 1
∫ e dt = − k
o
1 '
0 c1 k11

Thus,

8
 ∞

∫ e dt 1
τ R1 kc' 1  k F 2 k12  τ R1 kc' 1
0
=− ×  − 1  λ = × β1 = f1 β1
∞
τ R' 1 kc1  k F 1k22  τ R' 1 kc1
∫ e dt
o
1
0

Similarly, we have:
∞

∫ e dt 2
τ R 2 kc' 2
0
= × β2 = f2 β2
∞
τ R' 2 kc 2
∫ e dt
o
2
0

Notice that the PI parameters in the interacting loops are used to be more conservative
than those in single loops. In another words,
f1 ≥ 1; f2 ≥ 1
The multi-loop control should be beneficial when the sum of absolute values of the

Remarks:
1. If λ is assumed not vary with frequency, and the process under study is FOPDT,
λ>1, f1 lies in the range 1< f1 <2, while 0.5<l<1, f1 lies in the range 1< f1 <3.
2. When f1 =1, β is equal to the ratio of response areas.
3. If β is small and f1 =is close to one, then the interacting control is favored for that
particular disturbance.
4. If β is large, the interacting control is un favorable for that particular disturbance.

The Relative Gain for Non-square Multivariable Systems

(J.C. Chang and C.C. Yu, CES Vol.45, pp. 1309-1323 1990)

Consider a non-square MV system.

ym×1 ( s ) = Gm×n ( s )un×1 ( s )
Define Moore-Penrose pseudo-inverse of the matrix G ( s) as:

G + ( s) = ( GT G ) GT ( s)
−1

Then, under close-loop control, the steady-state control input will be:
 ∂ui  +
u = G + (0) y d and   = gij (0) .
∂y
 j CL
Thus, the non-square relative gain is defined similarly to the square RGA, that is:

9
 −1
 ∂y   ∂yi   + T
Λ% =  i     = G (0) ⊗ G (0) 

 ∂u j OL  ∂u j CL 

Properties of the non-square RGA

1. Row sum of Λ% :
T
 n n n

RS = [ rs (1), rs (2), L , rs (m) ] =  ∑ λ%1 j , ∑ λ% 2j , L , ∑ λ% mj , ;
 j =1 j =1 j =1 

Where, rs(i) = G (0)G + (0)  ii

T
 n n n

CS = [ cs (1), cs (2), L , cs (n)] =  ∑ λ% j1 , ∑ λ% j 2 , L , ∑ λ% jn , = [1, 1, L, 1]
T
2.  j =1 j =1 j =1 

Where, cs (i ) = G + G (0)  ii ; (Note: G + G = (GT G ) −1 GT G = I )

3. 0 ≤ rs(i) ≤ 1, ∀i = 1, 2,L, m
m n
4. ∑ rs(i) = ∑ cs(i) = n
i =1 j =1

m m n n m n
Note: ∑ rs(i) = ∑∑ λ% = ∑∑ λ% = ∑ cs ( j ) = n
i =1 i =1 j =1
ij
j =1 i =1
ij
j =1

5. Non-square RGA is invariant under input scaling, but is variant under output
scaling:

( GS ) ⊗ (GS )+  = (G ⊗ G+ ) ( SG ) ⊗ ( SG )+  ≠ (G ⊗ G+ )
T T T T

6. Let P1 and P2 are permutation matrices. Then, Λ% ( PGP

1
%
2 ) = P1 Λ (G ) P2

A. Multi-loop BLT-Tuning:

I. BLT-1 method:

a. Calculate the Ziegler-Nichol settings for each PI controller by using the

diagonal element of G, i.e. gi,i.
b. Assume a detuning factor “F”, and calculate controller settings for loops.

kc ,i = k ZN ,i / F ; τ R ,i = (τ R ,i )ZN F

c. Define: W(iω ) = −1 + det  I + G( iω )Gc ( iω ) 

10
 d. Calculate the closed-loop function Lc(iω):

W( iω )
Lc (iω ) = 20 log
1 + W( iω )

e. Calculate the detuning factor F until the peak in the Lc log modulus curve is
equal to 2N, that is:

 W( iω ) 
Lcm = Max 20 log  = 2N
ω
 1 + W( iω ) 

II. BLT-2

a. Find BLT-1 PI controllers.

b. Choose a second detuning factor FD. FD should be greater than one.
c. Compute τD,j as:

τ D, j =
(τ )
D , j ZN

FD

d. Calculate W(iω) and Lc(iω).

e. Change FD until Lmax

C is minimized, maintaining FD > 1. The trivial case may

C is minimized for FD = ∞ , i.e., no derivative action.

result where Lmax

f. Reduce F in the P and I modes, until Lmax

C = 2N .

III. BLT-3

The objective is to estimate the level of imbalance in detuning the BLT-1

controller and compensate for it.
Consider the PI controller:

 1
t

u j = u j (0) + kC , j  e j +
 ∫
τ R, j 0
e j dt  ;

u j (0) = 0
 
At steady state,

11
 ∞
kC , j
lim u j (t )  =
t →∞ τ R, j ∫ e (t )dt
0
j

So,
∞
τ R , j u j (∞)
∫ e (t )dt =
0
j
kC , j

Notice that:
u (∞) = G −1 (0) R − G −1GL (0)d (∞)

For unit step set-point input:

u j (∞) = G −1 (0)[0,..., 0,1, 0,..., 0]T

= [ g i , j (0); i, j = 1,..., N ][0,..., 0,1, 0,..., 0]T

For unit step load disturbance:

N
ui (∞) = i th row of G −1 (0)GL (0) = ∑ [ gi , j (0) g L, j (0)]
j =1

Then, ITEj becomes:

u j (∞)τ R , j
ITE j =
kC , j

Let,
N ITE j
Sj = ∑ + ITE j −load
i =1 N

τ R, j  N g j ,i (0) 
Sj = × ∑ + g j ,i (0) g L,i (0) 
kC , j  i =1 N 

Let S max = Max S j

j

Smax
Fj = F
Sj

The PI controller parameters becpme:

kc ,i = kZN ,i / Fj ; τ R ,i = (τ R ,i ) ZN Fj

IV. BLT-4

12
 a. BLT-3 is used to get individual PI controllers as described above.
b. BLT-2 procedure is used with individual FD factors for each loop:

S max
FD , j = FD
Sj

V. Tyreus Load-Rejection Criterion (TLC)

The best variable pairing is the one that gives the smallest magnitudes for each
element of X,(i.e. Xi) of the following:

X ( iω ) = ([ I + GGC ]−1 GL L )
( iω )

VI. Summary

BLT-1--- PI, BLT-2---PID,

equal Fi equal Fi

BLT-3---PI, BLT-4-----PID,
unequal Fi unequal Fi

B. Parallel-design method---Modified Z-N methods for

TITO Processes

This method is based on A modified Z-N method for SISO control system. To derive
this modified Z-N method, ageneral formulation is to start with a given point of the
Nyquist curve of the process:
j ( −π +ϕ p )
G p ( jω ) = rp e (1)

And to find a regulator GR

13
   1 
GR ( jω ) = k  1 + j  ωτ D −  (2)
  jτ Rω  

To move this point to B = rs e j ( −π +ϕs ) (3)

An amplitude margin (i.e. gain margin) design corresponding to ϕ s = 0 and

1
rs = .
Am
A phase margin design corresponds to rs = 1 and ϕ s = ϕm
j ( −π +ϕ p +ϕ R )
From Eqs.(1)~Equ.(3), we have: rs e j ( −π +ϕs ) = rp rR e , so that

rs
rR = and ϕ R = ϕ s − ϕ p
rp
In other words,

  1  j (ϕ R )
GR ( jω ) = k  1 + j  ωτ D −   = rR e = rR cos ϕ R + jrR sin ϕ R
  jτ Rω  

Or,

 1 
k = rR cos ϕ R =
rs
rp
(
cos ϕ s − ϕ p ) and  ωτ D − (
 = tan ϕ s − ϕ p
τ Rω 
)

The gain is uniquely determined. Only one equation determines τ R and τ D .

Let τ D = ατ R , where α is often chosen as α ≈ 0.25 . Another method to specify α

is as follows:

0.413 g (0)
α= , where κ =
3.302κ + 1 g ( jωc )

 1 
From  ωτ D −
τ ω
−1
( )
 = tan ϕ s − ϕ p , τ D can be solved to obtain:
 R 

1 
τD = − tan(ϕ s − ϕ p ) + 4α + tan 2 (ϕ s − ϕ p )  and
2ω  
1
τR = τD
α

Consider a stable 2 × 2 process :

14
 y1 ( s)   g11 ( s) g12 ( s )   u1 ( s ) 
 y ( s )  =  g ( s ) g ( s )  u ( s ) 
 2   21 22  2 

 c1 ( s )  c1 ( s ) 0 
c ( s )  =  0 c2 ( s) 
 2  

c2 g12 g 21 g g
g1 = g11 − = g11 − −112 21
1 + c2 g 22 c2 + g 22
g12 g 21
g 2 = g 22 −
c1−1 + g11

Let

Ai = rai e j ( −π +ϕai ) = gi ( jωi )

Bi = rbi e j ( −π +ϕbi ) = gi ( jωi )ci ( jωi )

  1 
ci ( jω ) = k 1 + j  ωτ Di +   ; i = 1, 2
 jτ Riω  
 

Take PI controller as example.

ci ( jω ) = kci (1 − j tan(ϕbi − ϕ ai ) ) ; i = 1, 2

And, gi ( jωi )kci = cos(ϕbi − ϕai )rbi e j ( −π +ϕai )

15
 rai e j ( −π +ϕia ) ⋅ kci (1 − j tan(ϕbi − ϕai )) = rbi e j ( −π +ϕia )
⇓
rbi j (ϕai −ϕbi ) rbi r
e = cos(ϕai − ϕbi ) + j bi sin(ϕai − ϕbi ) = kci (1 − j tan(ϕbi − ϕ ai ))
rai rai rai
⇓
rbi
kci = cos(ϕai − ϕbi )
rai
rbi
kci ⋅ gi ( jω ) = cos(ϕ ai − ϕbi ) ⋅ rai e j ( −π +ϕia ) = rbi cos(ϕ ai − ϕbi ) ⋅ e j ( −π +ϕia )
rai

By setting i equal one and two, one will obtain two equations with kc1 and kc2 as
unknowns, and, thus, can be solved. But, there are very tedious procedures to find the
controller gains (such as:such kc1 and kc2) and frequency ω11 and ω22 that satisfy the
phase criteria. (see the reference: I&EC Res. 1998, 37, 4725-4733, Q-G Wang, T-H
Lee, and Y. Zhang)

C. Independent design method

---IMC Multi-loop PID Controller

16
 f1 [(g11) -]-1
_
G
_ f2 [(g22)-]-1

g11 _

g22 _

GC ,i = ( Gi ,i )− f i ; i = 1,..., n
−1

The stability is guaranteed for any stable IMC filter that satisfies either of the
following:

g i ,i (iω )
f i (iω ) < f R*,i (iω ) = ; i = 1, 2,..., n
∑
j , j ≠i
gi , j (iω )

g i ,i (iω )
f i (iω ) < f C*,i (iω ) = ; i = 1, 2,..., n
∑
j , j ≠i
g j ,i (iω )

Imc Row interaction measure [Economou and Morari]

1
∑
j , j ≠i
g i , j (iω )
Ri (iω ) = = ; 0 ≤ω ≤ ∞
1 + f R*,i (iω ) ∑g j
i, j (iω )

1
∑
j , j ≠i
g j ,i (iω )
Ci (iω ) = = ; 0 ≤ω ≤ ∞
1 + fC*,i (iω ) ∑g
j
j ,i (iω )

For significant interaction: 0.5 ≤ Ri , Ci ≤ 1 ⇒ f * <1

17
 For small interaction: 0.0 ≤ Ri , Ci ≤ 0.5 ⇒ f * >1

D. Chien-Huang-Yang’s multi-loop PID---with no

proportional and derivative kicks

1. Controllers for SISO loop:

 1 
Controller: u ( s ) = kC  − y ( s ) + [r ( s ) − y ( s)] − τ D sy ( s) 
 τ Rs 

y k /(τ s )G p
= C R
r 1 + kC /(τ R s )G p

a. Time constant dominant processes:

Re − Ls
GP = ; R = slope of the initial unit step response
s
Re− Ls R(1 − Ls)
GP = ≈
s s
y 1 − Ls 1 − Ls
= ≈ 2 2
r  τR  τ C s + 1.414τ C s + 1
 − τ R L  s 2 + (τ R − L ) s + 1
 RkC 
(1.414τ C + L)
⇒ kC = ; τ R = 1.414τ C + L
R(τ + 1.414τ C L + L2 )
2
C

b. Deadtime dominant processes:

k P e− Ls k P (1 − Ls )
GP = ≈
τ s +1 τ s +1

18
 y 1 − Ls
=
r  τ Rτ   τ 
 − τ R L  s2 +  R + τ R − L  s + 1
 kC k P   kC k P 
1 − Ls
≈
τ C s + 1.414τ C s + 1
2 2

1 −τ C2 + 1.414τ Cτ + Lτ
⇒ kC = ;
k P τ C2 + 1.414τ Cτ + L2
−τ C2 + 1.414τ Cτ + Lτ
⇒ τR =
τ +L

Derivation of the PID controller parameters is similar to the above PI

derivations except that the deadtime approximation:

1 − 0.5 Ls
e − Ls ≈
1 + 0.5 Ls

19
20
21
 2. Controllers for multi-loop system

 y  k k  g1,1
At ω → 0 ;   = g1,1 1 − 1,2 2,1  =
 u1 loop 2 closed  k1,1k2,2  RGA(λ )

 y
At ω → ∞ ;   = g1,1
 u1 loop 2 closed

a. For RGA>1, multi-loop controller tuning based on the process

model in the main loop should provide satisfactory closed loop
results. This is because:
b. For RGA < 1,

kC ,i = ( kC )based on main loop RGA(λi ,i )

(τ )
R ,i based on main loop
τ R ,i =
RGA(λi ,i )

τ D ,i = (τ D ,i )based on main loop RGA(λi ,i )

The closed-loop time constant is chosen according to the value of L/τ

in three different ranges, that is: L/τ < 0.2, 0.2 < L/τ < 0.5, and L/τ >
0.5.
For details, see the original paper.
.
IX. Robustness of Closed-loop System.

The final pairing and the controller tuning is checked for robustness by
plotting DSO and DSI as functions of frequency, [Doyle and Stein]. The
singular values below 0.3-0.2 indicate a lack of stability robustness.

DSO(iω ) = σ [ I + ( GGC ) ]( iω )
−1

DSI ( iω ) = σ [ I + ( GC G ) ]( iω )
−1

22
E. Design Method based on Passivity

1. Hardware simplicity and relative effortlessness to achieve failure tolerant design,

multi-loop control is the most widely used strategy in the industrial process
control.
2. Current multi-loop control design approaches can be classified into three
categories: detuning methods (Luyben, 1986), independent design methods
(Skogestard and Morari, 1989), and sequential design methods (Mayne, Chiu and
Arkun, 1992).
3. Loop interactions have to be taken into considerations, as they may have
deteriorating effects on both control performance and closed-loop stability.
4. It is desirable if the multi-lop control system is decentralized unconditionally
stable (i.e., any subset of the control loops can be independently to an arbitrary
degree or even turned off without endangering close-loop stability.
5. Independent design is based on the basis of the paired transfer function while
satisfying some stability constraints due to process interactions.
6. Perhaps the mostwidely used decentralized stability conditions are those
µ-interaction measure.
7. Passivity Concept:
The rate of change of the stored energy in the tank is less than the power supplied
to it.
Inlet Flowrate
Fi

Fo
h Outlet Flowrate

Potential energy stored in the tank: S (h) = 12 Ahρ gh = 12 Aρ gh 2

Increment of potential energy per unit time: w(t ) = ρ Fi (t ) gh(t )

The rate of change of the storage function:
dS
= −Cv ρ gh h + ρ gFi h = −Cv ρ gh h + w < w ∀h > 0
dt
The rate of change of the stored energy in the tank is less than the power
supplied to it. Therefore this process is said to be strictly passive.

23
Passive(Willems 1972): if a non-negative storage function S(x) can be found s.t.:
t
S(0)=0 and S ( x) − S ( x 0 ) ≤ ∫ yT (τ )u (τ )dτ for all t>t0≥0, x0, x∈ X, u∈ U.
t0

t
Strictly passive: if S ( x) − S ( x 0 ) < ∫ yT (τ )u (τ )dτ
t0

Where, y is the output of a system, u is the input to the system.

 KYP Lemma
 Nonlinear control affine systems (Hill & Moylan 1976)
x& = f ( x) + g ( x)u
y = h( x )
where x ∈ X ⊂ R n , u ∈ U ⊂ R m , y ∈ Y ⊂ R m
The process is passive if
∂S T ( x )
Lf S ( x) = f ( x ) ≤ 0,
∂x
∂S T ( x )
Lg S ( x ) = g ( x ) = hT ( x )
∂x
 KYP Lemma
A linear system (Willems 1972) G(s):=(A,B,C,D) is passive if there exists a
positive definite matrix P such that:

 AT P + PA PB − C T 
 T ≤0
 B P − C − D − D 
T

The system is strictly passive if

 AT P + PA PB − C T 
 T <0
 B P − C − D − D 
T

Definition:
An LTI system S: G(s) is passive if :
(1) G(s) is analytic in Re(s)>0;
(2) G(jw)+G*(jw)≥0 for all that jw is not a pole of G(s);
(3) If there are poles of G(s) on the imaginary axis, they are non-repeated and the
residue matrices at the poles are Hermitian and positive semi-definite.
G(s) is strictly passive if:
(1) G(s) is analytic in Re(s) ≥ 0;
(2) G(jw)+G*(jw)>0 ∀ω ∈ (−∞, ∞) .

24
Theorem 1: For a given stable non-passive process with a transfer function matrix
G(s), there exists a diagonal, stable, and passive transfer function matrix
W(s)=w(s)I such that H(s)=G(s)+W(s) is passive.
[Proof]:

λmin ( H ( jω ) + H * ( jω )) = λmin (G ( jω ) + G* ( jω ) + (W ( jω ) + W * ( jω ))

Since both (G+G*) and (W+W*) are Hermitian, from the Weyl inequality, we
have:

λmin ( H ( jω ) + H * ( jω )) ≥ λmin (G ( jω ) + G* ( jω )) + λmin (W ( jω ) + W * ( jω ))

=λmin (G ( jω ) + G* ( jω )) + 2 Re(W ( jω ))

Thus, if:
1
Re(W ( jω )) ≥ λmin (G ( jω ) + G* ( jω ))
2
H(s) can be render passive. On the other hand, if
1
Re(W ( jω )) > λmin (G ( jω ) + G* ( jω ))
2
H(s) will be strictly passive.

Properties of Passive Systems:

 A passive system is minimum phase. The phase of a linear process is within
[-90º, 90º]
 Passive systems are Lyapunov stable
 A passive system is of relative degree < 2
 Passive systems can have infinite gain (e.g., 1/s)

Passivity Theorem :

If G1 is strictly passive and G2 is passive, then the closed-loop system is L2

stable.

 A strictly passive process can be stabilized by any passive controller

25
 (including multi-loop PID controllers) even if it is highly nonlinear and/or
highly coupled
 Control design based on passivity
 Excess or shortage of passivity of a process can be used to analyse whether
this process can be easily controlled
 Passivity based controllability study

A non-passive process can be made passive using feedforward and/or feedback

passification:

Gff Gfb

G G

The excess or shortage of passivity can be quantified using:

 Input Feedforward Passivity (IFP) (Sepulchre et al 1997) - If a system
G with a negative feedforward of νI is passive, then G has excessive
input feedforward passivity, i.e., G is IFP(ν).

 Output Feedback Passivity (OFP) (Sepulchre et al 1997) - If a system

G with a positive feedback of ρI is passive, then G has excessive
output feedback passivity, i.e., G is OFP (ρ).

Agin, use the following figure:

If G1 is IFP(ν) and G2 is OFP(ρ), then the closed-loop system is stable if ρ+ν>0.

In other words, a processs shortage of passivity can be compensated by another
process’s excess of passivity.

 Passivity Index
The excessive IFP of a system G(s) can be quantified by a frequency dependent

26
passivity index
∆
1 
ν F [G ( s ), ω ]=λmin  [G ( jω ) + G *( jω )] 
2 
Assume the true process is GT ( s ) = G ( s ) + ∆( s)
The passivity index of the true process can be estimated as
1 1 
ν (GT ( jω )) = −λmin   ∆( jω ) + ∆* ( jω )  + G ( jω ) + G* ( jω )  
2 2 
1  1 
≤ −λmin   ∆( jω ) + ∆* ( jω )   − −λmin  G ( jω ) + G* ( jω )  
2  2 
=ν (G ( jω )) +ν (∆ ( jω ))

Properties of the Passivity Index

1. Comprises gain & phase information of the uncertainty

j
Maximum gain

Passivity index

σ
∆(σ)
 

2. Always no greater than the maximum singular value.

ν F [∆( s), ω ] ≤ σ max [∆( jω )] for any ω ∈ R

27
 Passivity Theorem 2: If the multivariable process is strictly passive, then the
closed-loop system is stable if the multi-loop controller is passive.

Theorem 1: A closed-loop system comprising a stable subsystem G(s) and a

decentralized controller K(s)=diag(ki(s)), w(s) is a stable and minmum phase, and

ν (W ( jω )) < −ν (G + ( jω ))

Then the closed-loop system will be decentralized unconditional stable, if

K(s)=diag{ ki' ( s ) } is passive, where,

ki' ( s ) = ki+ [1 − w( s)ki+ ( s)]−1 and ki+ = U ii ki

−1
K ' ( s ) = U −1K ( s)  I − w( s )U −1K ( s) 

Notice that the above figure is equivalent to the one in the following:

Rescaling of the system

Let D(s) be a diagonal, real and constant scaling matrix.

The scaling matrix D(s) is to make

ν ( D −1G + D ( jω )) < ν (G + ( jω ))

and

28
 +
D −1G + (0) D + D −1 G + (0)  D > 0

Design procedures:

1. Find matrix U and calculate G + ( s) .

2. Check the pairing. Examine the proposed pairing using DIC condition:
T
G + (0) M + M G + (0)  > 0

3. Use matrix M obtained in the step 2 to derive D, D = M 1/ 2

4. Calculate ν ( D −1G + ( jω ) D ) for different frequency points. These frequency

points form a set Ω ∈ [ 0, ωE ] where ωE is the frequency which is high enough

sych tant ν ( D −1G + ( jω ) D) → 0 for ω > ωE .

5. For each loop of the controller, solve problem:

min(−γ i )
kci ,τ Ri

such that

1 γi
<1
 1  jω
1 + Gii+ ( jω )kc+,i 1 + 
 jτ R ,iω 

and

kc+,iν s (ω )
τ R2 ,i ≥ , ∀ω ∈ R, i = 1,L , n
1 − kc+,iν s (ω )  ω 2
 

6. Obtain the final controller settings: kc ,i = U ii kc+,i

This method is limited to open-loop stable processes.

29
 ν F (∆( s ), ω ) ≥ −ν F (W ( s ), ω ), ∀ω ∈ R

Robust Stability Condition

If the uncertainty is passive, then the controller is only required to render system T
strictly passive to achieve robust stability even if ∆ is very large.

If the uncertainty’s passivity index is bounded by

ν F ( ∆( s ), ω ) ≥ −ν F (W ( s), ω ) , ∀ω ∈ R

where W(s) is minimum phase, the closed-loop system will be robust stable if
system

T ( s )[ I − W ( s )T ( s)]−1

is strictly passive.

The basic idea:

1. Characterise the uncertainty in terms of passivity using IFP or OFP.
2. Derive the robust stability condition for systems with uncertainties bounded by
their passivity indices.
3. Develop a systematic procedure to design the robust controller which satisfies
the above stability condition.

Passivity Based Robust Control Design

 Blended approach
 Design a controller that satisfies the small gain condition at high
frequencies and satisfies the passivity condition at low frequencies
(Bao, Lee et al 1998)
 Based on the bilinear transformation

30
  Multi-objective control design
 Design a controller that satisfies the passivity condition for robust

stability and achieves H∞ control performance (Bao, Lee et al 2000,

2003)
 Based on KYP lemma and Semi-Definite Programming

Example:

 0.126e −6 s − 0.101e −12 s 

 
G ( s) =  60s +−18 s (48s + 1)(45s + 1) 
−8 s
 0.094e − 0.12e 
 38s + 1 35s + 1 

Passivity index

0.0.0404

P
0.0.0202
as
0.0
si 0.0 0

vit
-0.0
-0.-0.020

-0-0.03.0
110-40
- 10-2
1 1 100 10+2
1 10+41 4
(rad/min

F. Design by Sequential Loop Closing

Advantages of sequential design:

1. Each step in the design procedure involves designing only one SISO controller.
2. Limited degree of failure tolerance is guaranteed: If stability has been achieved

31
 after the design of each loop, the system will remain stable if loop fail or are taken
out of service in the reverse order of they were designed.
3. During startup, the system will be stable if the loops are brought into service in the
same order as they have been designed.
4.
Problems with sequential design:
1. The final controller design, and thus the control quality achieved, may depend on
the order in which the controllers in the individual loops are designed.
2. Only one output is usually considered at a time, and the closing of subsequent
loops may alter the response of previously designed loops, and thus make iteration
necessary.
3. The transfer function between input uk and output yk may contain RHP zeros that
do not corresponding to the RHP zeros of G(s).

Notations:

1. G(s): the n × n matrix of the plant, G ( s ) = {gij ( s ); i, j = 1,L , n}

2. C ( s) = diag{ci ( s); i = 1,L , n}

3. S = ( I + GC )−1 ; H = I − S = GC ( I + GC ) −1

4. G% = diag{gii ( s ); i = 1,L , n}

1
5. S% = diag{si ( s ); i = 1,L , n} = diag{ ; i = 1,L , n}
1 + gii ci
g c
6. H% = diag{hi ( s ); i = 1,L , n} = diag{ ii i ; i = 1,L , n}
1 + gii ci

% −1 = {γ ; i, j = 1,L , n}
7. Γ = GG ij

% −1G
8. CLDG = GG d

9. E = (G − G% )G% −1

G M C M
10. G =  k  ; C= k ;
 L O  L O

11. Sk = ( I + Gk Ck ) ; H k = Gk Ck ( I + Gk Ck )
−1 −1

Hk 0  Sk 0
12. Hˆ k =   ; Sˆk =  ; i = k + 1, K + 2, L N
 0
%
hi  0 s%i 

32
 S = ( I + GC )−1 = [ I + GC
% + (G − G% )C ]−1

{ )}
−1
( ) (
−1
=  I + (G − G% )C I + GC
%  I + GC
%
 

{ )}
−1
( )  (
−1 
=  I + (G − G% )G% −1GC
% I + GC
% %
I + GC


( ) ( I + EH% ) ( )
−1 −1 −1
%
= I + GC = S% I + EH%

Design procedures:
) ) )
( )
−1 −1
In each of the following step, S = S k ( I + Ek H k ) ; Ek = (G − Gˆ k ) Gk

Determine ci such that W p SWD is minimized.

Step 0. Initialization. Determine the order of loop closing by estimating the

required bandwidth in each loop. Also estimate the individual loop designs
in terms of H% .
Step 1. Design of controller c1 by considering output 1 only. In this case, we have

Gˆ k = G% k and Hˆ k = H%

Step k. Design of controller ck by consider outputs 1 to k. Here,

Gˆ k = diag{G% k , gii }; i = k + 1, k + 2,L , n and

Hˆ k = daig{H k −1, h%i } ; i = k , k + 1,L , n

33
Sequential Design Using Relay feedback Tests of Shen and Yu

The relay feedback system for SISO auto-tuning is as shown in the follwing figure:

When constant cycles appear after the system has been activated, the ultimate gain
and ultimate frequency of the open-loop system can be approximated by measuring
the magnitude and period (see the following figure) and by the following equations:

4h 2π
Ku = ; ωu =
πa Pu
The Z-N tuning method can be used to determine the controller parameters:

PI Controller: K c = 0.45K u , τ R = Pu /1.2,

PID Controller: K c = 0.60 Ku , τ R = Pu /1.2, τ D = 1.25 Pu
Or, use the Tyreus-Luyben’s formula to give more conservative response:
PI Controller: K c = K u / 3.2, τ R = 2.2 Pu ,
PID Controller: K c = K u / 2.2, τ R = 2.2 Pu , τ D = Pu / 6.3
To avoid the difficult mathematics envolved in the formulation of sequential
design, Shen and Yu suggested to use the relay-feedback test as shown in the
following figure:

34
 The controller for a 2 × 2 system is suggested:

PI Controller: K c = K c, ZN / 3, τ R = 2 Pu

Analysis:
The sequential design is derived by considering the multi-loop control system as
coupled SISO loops. For a 2 × 2 system as example, the equivalent SISO loops are:
1
g1 ( s) = g1,1 ( s ) {1 − (1 − ) h ( s)}
λ (s) 2
1
g 2 ( s) = g 2,2 ( s ) {1 − (1 − ) h ( s)}
λ (s) 1

g C,i gi ,i
Where, hi (s) = ; i = 1, 2
1 + gC ,i gi ,i

Notice that, if there is damping in g1 or g 2 , this damping should come from either
h1 or h2 . According to tis study, a closed system having an FOPDT process and a
modified ZN tuned PI controller will result in a closed-loop system (i.e. h1 and h2 )
having damping factor greater than 0.6. It is thus postulate that the open-loop transfer
functions g1 ( s ) and g 2 ( s ) can be approximated by:

kp τ p 2 s + 1 −θ s
G ( s) = ⋅ ⋅e
τ 2 s 2 + 2τζ s + 1 τ p1s + 1

Then, the stability region of the equivalent SISO loops are explored with the

parameters: τ p1 ,τ p = 0 ~ 10, k p = 1, τ = 5, ζ = 0.1 ~ 1, θ / τ = 0.02 ~ 0.2 . The results

35
are given in the following figure. It can be seen that the modified ZN tuning formula
proposed greatly improve the stability.

On the other hand, the convergence of the sequential design for the multi-loop
controller is formulated as the problem of finding the roots of simultaneous algebraic
equation using sequential iterations.
The simultaneous equations are obtained from the conditions of phase crossover for
the two loops, that is:
Im  g1 ( jωu ,1 , jωu ,2 ) 
F1 ( jωu ,1 , jωu ,2 ) = tan −1 = −π
Re  g1 ( jωu ,1 , jωu ,2 ) 
Im  g 2 ( jωu ,1 , jωu ,2 ) 
F2 ( jωu ,1 , jωu ,2 ) = tan −1 = −π
Re  g 2 ( jωu ,1 , jωu ,2 ) 

The convergence of the iteration is guaranteed by a sufficient condition of the

following:
 ∂F1   ∂F2 
   
 ∂ωu ,2 ωu ,1  ∂ωu ,1 ωu ,2
<1
 ∂F1   ∂F2 
   
 ∂ωu ,1 ωu ,2  ∂ωu ,2 ωu ,1

36
 The procedures of this proposed sequential design are summarized with the flow
chart as shown.

37