Download as pdf or txt
Download as pdf or txt
You are on page 1of 6

2008 10th Intl. Conf.

on Control, Automation, Robotics and Vision


Hanoi, Vietnam, 17–20 December 2008

Robust Control for Axial-Flow Compressor – An


Algorithm
Tiebao Yang1 , Chun-Yi Su2 , Xiang Chen1 , and Henry Hu3
1
Department of Electrical & Computer Engineering, University of Windsor, Windsor, Ontario N9B 3P4, Canada
2
Department of Mechanical & Industrial Engineering, Concordia University, Montreal, Quebec H3G 1M8, Canada
3
Department of Mechanical, Automotive & Materials Engineering, University of Windsor, Windsor, Ontario N9B 3P4, Canada
E-mail: xchen@uwindsor.ca (X. Chen)

Abstract—Local L2 gain method has been pursued to analyze is known [17] that the key issue for a successful L2 gain
the robustness of a family of stabilizing controllers near the rotat- analysis and design is whether one could find a Lyapunov
ing stall point in axial flow compressors. An matrix inequality is
function to solve the famous Hamilton-Jacobi-Bellman (HJB)
involved in computation of L2 gain which is usually very difficult
to solve without resorting to numerical method. In order to reduce equation or inequality. Therefore, in principle, if one could
the computational cost, an algorithm is proposed in this paper to obtain a class of stabilizing controllers and a feasible Lyapunov
greatly simplify the computational complexity by converting the function candidate, then the L2 gain, which is an indication
matrix inequality into its quadratic form. Rather than obtaining of robustness in the norm bound form, could be calculated
accurate approximation of the local L2 gain, the results would
out for every stabilizing controller and hence the controller
facilitate the robust design of the axial flow compressor control.
Index Terms—Robustness; bifurcation stabilization; L2 gain; achieving the smallest L2 gain would be deemed as the most
compressor control. robust one in that given class. For bifurcation systems, since a
parameter is involved, at the first glimpse, one would think that
I. INTRODUCTION a Lyapunov function to facilitate the L2 gain analysis should be
Rotating stall and surge are aerodynamic instabilities in- also parameterized which may cause extra analytic complexity.
troduced by nonlinear bifurcations of the operating point in However, when one takes a closer look at the stabilized bifur-
axial flow compressors [15]. The engine performance could cation system [9], even if there still exists equilibrium point
be adversely affected since compressors are the vital parts of bifurcation, the system behavior will not change significantly
turbine-based jet engines. In the past, families of linear or around the critical position. Observation of this fact motivates
nonlinear feedback controllers have been derived to address the us to conduct local L2 gain analysis for bifurcation stabilizing
rotating stall problem in order to extend the stable operating controllers when applied to the corresponding critical system
range and hence improve engine performance [3], [6], [8], and to predict that the L2 performance of the stabilizing
[9], [19], [22]. Regarding these progresses, an interesting controller at the critical situation would reasonably indicate
question raised further is: to what extent, can these stabilizing similarity with that at bifurcated situation.
It has been shown [23] that an matrix inequality is involved
control designs be deemed as robust? In [4], [5], local robust
in computation of L2 gain which is usually very difficult to
analysis methods are presented based on the projection method
solve without resorting to numerical method. If our ultimate
[10] for a class of stationary and Hopf bifurcation stabilizing
goal is to find the most robust controller rather than accurately
controllers [9]. The results show that the local admissible
estimate the L2 gain, we can convert the matrix inequality
uncertainty set can be characterized using the coefficients in
into its quadratic form, which would greatly simplify the
the Taylor series expansion. It is pointed out that, since the
computational complexity. In this paper, we shall present
characterizing conditions depend explicitly on the stabilizing
an algorithm that facilitates the robust design of the axial
control gain, one can numerically compute and compare the
flow compressor control. It should be noted that the method
‘size’ of the uncertainty set that can be tolerated by each con-
could be generalized (not very trivially) to other bifurcation
troller, which, in other words, provides an effective approach to
stabilizations.
conduct the so-called ‘robust design’. The reason that we use
the ‘robust analysis’ approach to conduct the ‘robust design’ II. P RELIMINARY T HEORIES
is due to the fact that, for a nonlinear system, one usually does The system under consideration is the following nonlinear
not know how to characterize all stabilizing controllers but just control system F with bifurcation parameter γ, subject to
a subset of them. uncertainty Δ as a smooth mapping as shown in Fig. 1:
In our previous work [23], L2 gain method is applied to
characterize the robustness of stationary bifurcation stabiliza- ẋ = f (γ, x) + g(x)u + bw
(1)
tion. It has been shown that the advantage provided by the z = cx
L2 approach is that the robustness can be explicitly measured where γ is the bifurcation parameter with a critical value 0;
by the L2 gain values achieved by different controllers. It x∈Rn ; b and c are real valued vectors; z ∈ Rq is the desired
978-1-4244-2287-6/08/$25.00 
c 2008 IEEE 245 ICARCV 2008
= L∗0 x + Q∗0 [x, x] + C0∗ [x, x, x] + · · · + bw + · · ·

z w where L∗0 , Q∗0 [x, x], and C0∗ [x, x, x] are vector valued linear,
quadratic, and cubic terms of f ∗ (x), respectively; and f ∗ (x) =
F f (x) + g(x)Kx.
y u For Lyapunov functions, the result in [7] is applied, where
C algorithms have been developed to construct a family of
Lyapunov functions for nonlinear systems in critical cases. The
Fig. 1. Nonlinear Control System in Robust Consideration
Lyapunov functions have the form:
 
V (x) = α xT P x + ρ[x, x, x] , (3)

output signal for performance evaluation; u is a scalar feedback where α is a positive number; and ρ[x, x, x] is the cubic term in
control; w ∈ R is a scalar disturbance signal; f (·, ·) : R × x. It has been shown [23] that only quadratic term contributes
Rn →Rn and g(·) : Rn →Rn are all smooth functions with to the analysis of local L2 gain of the bifurcated systems, and
f (γ, 0) = 0. In this paper, the system (1) is assumed to satisfy: the cubic term can be ignored. Let ∗ and r∗ be a left row and
 a right column eigenvector of L∗0 corresponding to the critical
df (γ, x) 
(i) L(γ) = has a simple eigenvalue (critical eigenvalue such that ∗ r∗ = 1, a simplified algorithm to obtain
dx x=0
mode) λ(γ) satisfying: λ(0) = 0 and λ (0) = 0, while all the matrix P is listed as follows.
other eigenvalues are stable in a neighborhood of γ = 0. Algorithm 2.3: [7] (Construction of Lyapunov functions)
(ii) The critical mode of L(0) is not observable, nor linearly
1) Compute ∗ and r∗ , choose a basis {r∗2 , r∗3 , ..., r∗n }
controllable by u, nor affectable by w, i.e., for all  and
for stable subspace E s , which is defined as the eigen-
r, if L(0) = 0 and L(0)r = 0, we have g(0) = 0 and
vectors (and the generalized eigenvectors, if any) of
b = 0.
L∗0 corresponding to the stable eigenvalues of L∗0 (see
The corresponding critical system is then taken as γ = 0: [Guckenheimer & Holmes, 1983; Troger & Steindl,
ẋ = f (0, x) + g(x)u + bw = f (x) + g(x)u + bw 1991]). Compute the dual basis {∗1 , ∗2 , ..., ∗n } to the
z = cx
(2) basis {r∗1 , r∗2 , ..., r∗n }, where ∗n = ∗ and r∗n = r∗ .
Compute the coefficients
Definition 2.1: Assume u = 0 and w = 0. The nonlinear β1 = ∗ Q∗0 [r∗ , r∗ ]
system in (1) is said to have local bifurcation stability if the   
origin is locally asymptotically stable in a sufficiently small β2 = 2∗ C0∗ [r∗ , r∗ , r∗ ] − 2Q∗0 r∗ , L∗− ∗ ∗ ∗
0 Q0 [r , r ]
neighborhood around the origin and a small neighborhood of where
γ = 0.  L∗− ∗T ∗ ∗T ∗ −1 ∗T
0 = (L0 L0 +   ) L0
In [14], the L2 gain of an input-output mapping and the Check if β1 = 0 and β2 < 0. If not, we need to re-design
applications to robust nonlinear control have been discussed in a family of stabilizing bifurcation controllers.
details. By Small Gain Theorem [14], the L2 gain, measuring 2) Choose any real symmetric n×n matrix Π satisfying,
z2
w2 , can be actually interpreted as the measurement of control for all a∈E s and a=0: i) Πr∗ = 0; ii) aT Πa > 0; iii)
robustness when the uncertainty signal w is in norm-bounded aT (L∗T ∗
0 Π+ΠL0 )a < 0; and iv) β2 +Δs (Π) < 0, where
form.  ∗− T ∗−
 ∗ ∗ ∗
Δs (Π) = −Q∗T ∗ ∗
0 [r , r ] (L0 ) Π + ΠL0 Q0 [r , r ]
Definition 2.2: A feedback control u(x) is called local
stabilizing bifurcation control if, for w = 0, the closed-loop 3) Set P = ∗T ∗ + Π. 
nonlinear system in (1) ẋ = f (γ, x) + g(x)u(x) has local
Secondly, once we have obtained a family of stabilizing
bifurcation stability. The local L2 gain achieved by u(x) with
controllers and Lyapunov functions, the following theorem
respect to a norm-bounded uncertainty signal w is defined as
can be applied to calculate the local L2 gain achieved by the
the local L2 gain achieved by u(x) when applied to the critical
controllers u.
system (2). 
Theorem 2.4: [23] Given a class of local stabilizing bifur-
Firstly, in order to conduct the L2 gain analysis, we need
cation controllers u = Kx for the bifurcated system (2). Let
to synthesize a class of stabilizing controller and construct a
 
family of Lyapunov functions for system (2). For presentation V (x) = α xT P x + ρ[x, x, x]
simplicity, only local state feedback stabilizing bifurcation
controllers u = Kx [2], [9] will be considered to illustrate be the Lyapunov function constructed from Algorithm 2.3 for
the ideas in this paper. The Taylor expansion of system (2) the system (2). If there is an η > 0 such that
can be expressed as 2α2 1
ẋ = f ∗ (x) + bw M = α(L∗T ∗
0 Π + ΠL0 ) + ΠbbT Π + cT c ≤ 0,
η2 2

246
then there exists a small neighborhood D ⊂ Rn of the origin 1) Synthesize a class of local stabilizing bifurcation con-
such that, ∀x(0) ∈ D, the local L2 gain achieved by u will be trollers u for the system (1) with w = 0 by using the
no greater than η as a function of α. Furthermore, the smallest theorems in [2], [9];
possible η can be obtained through a search conducted on 2) Construct Lyapunov function V (x) for the corresponding
varying α.  critical system [7], and let αV (x) with α > 0 be the Lya-
punov function candidate to solve the HJB inequality;
This theorem guarantees that, for any disturbance signal 3) Apply the Algorithm 2.5 to find an upper bound of

w
2 < η1
z
2 + δ1 for some δ1 > 0, the perturbed nonlinear the local L2 gains η for all the stabilizing bifurcation
system will remain stable and, besides,
z
2 ≤ η
w
2 + δ2 controllers u.
for some δ2 > 0. We have the following definition for the L2 4) Find the smallest upper bound of the local L2 gain, say
gain of a stabilizing bifurcation control. η0 and its corresponding controller u0 .
Finally, in practice, the value of η can be computed by
Then the u0 will be the ‘most robust’ controller among the
converting M into a quadratic form through choice of a special
given class achieving the best performance. As stated before,
vector q. While this approach does give an explicit expression
in general, one cannot tell how to synthesize all stabilizing
of η in terms of system parameters, it is unfortunately q
bifurcation control. Therefore, while a general conclusion
dependent and hence may only give an estimation of η. A
could not be reached for the L2 -gain control of the nonlinear
more accurate way is to choose qiT = qi1 · · · qin as
bifurcated systems, it would be interesting to look at a sub-
an eigenvector of L∗0 corresponding to its stable eigenvalues
optimal problem, that is, conducting local L2 gain analysis for
λi , i = 1, · · · , n − 1. Then we have
an available subset of all stabilizing bifurcation controllers.
2αi2 T
q̄iT M qi = αi (q̄iT L∗T T ∗
0 Πqi + q̄i ΠL0 qi ) + q̄ ΠbbT Πqi III. L OCAL L2 G AIN A NALYSIS OF A XIAL F LOW
η̂i2 i C OMPRESSOR C ONTROL
1 Let us consider the local L2 gain problem for an axial
+ q̄iT cT cqi < 0.
2 compressor system based on the well-known Moore-Greizer
So we have model [15], [16]. In this section, a local feedback controller
is first synthesized for the unstable equilibrium point, then the
2αi2 T 1 method stated in Sec. 2 is applied to approximate the local L2
αi (λi + λ̄i )q̄iT Πqi + 2 q̄i1 ΠbbT Πqi + q̄iT cT cqi < 0,
η̂i 2 gain for the controlled system, and the most robust controller,
or which has the least L2 gain, is obtained.
2α2i 1
αi Li + Bi + Ci < 0, (4) A. Mathematical model of an axial compressors
η̂i2 2
The Moore-Greizer model, which describes the post-stall
where Li = (λi + λ̄i )q̄iT Πqi , Bi = q̄iT ΠbbT Πqi , Ci = dynamics in axial flow compressor, can be converted to the
q̄iT cT cqi , i = 2, 3, ..., n. The following procedures can be simpler form [3], [9], [15]:
stated to calculate an estimation of η:
Ψ̇ = β12 (Φ − Φγ (Ψ)),
Algorithm 2.5: Estimation of η.
Φ̇ = −Ψ + Ψc (Φ) + 6c3 ΦA2 , (7)
1) The minimum value of ηi can be achieved when taking Ȧ = 0.5σA(1 − Φ2 − A2 )
αi = − C i
Li , and the minimum value of ηi is given by
√ where A is the disturbance amplitude; Φ is the local flow
2 Bi Ci coefficient; Ψ is the pressure rise; γ > 0 is the throttle
η̂i = −
Li position; β, σ, c0 , c1 and c3 are all design constants with
β = 2.88, σ = 2.2867, c0 = 8/3, c1 = 1.5, c3 = −0.5; and
Clearly, η̂i > 0, since all λi s are stable eigenvalues. √
2) Find η̂ = mini (η̂i ).  Φγ (Ψ) = γ Ψ − 1, Ψc (Φ) = c0 + c1 Φ + c3 Φ3 .

It should be mentioned that the purpose of the above algorithm Define R as the amplitude square of the disturbance flow,
is only to provide the estimation of L2 gain that enables us i.e. R = A2 , then it is obvious that the system has two sets of
to calculate the most robust controller in a simple and fast equilibrium points [9]:
Set 1:
way. Also note that for all state feedback stabilizing bifurcation
controllers synthesized from theorems in [2], [9], the L2 gain Re = 0
could be estimated by applying Algorithm 2.5. Hence, the best Ψe = Ψc (Φe ) = c0 + c1 Φe + c3 Φ3e
2
robust controller could be determined by comparing all of these γ 2 = (1+Φ e)
Ψc (Φe )
L2 gain values and the controller corresponding to the smallest
Set 2:
value will be the one.
In summary, the general procedures to find the most robust Re = 1 − Φ2e , and Re =0
controller from a given class of local stabilizing bifurcation Ψe = Ψs (Φe ) = c0 + (c1 + 6c3 )Φe − 5c3 Φ3e
2
control u can be summarized as follows: γ 2 = (1+Φ e)
Ψs (Φe )

247
⎡ 1
√ ⎤
β 2 (x2
− (∇γ + γc ) Ψc 1 + x1 /Ψc + 2)
f (x) = ⎣ −x1 − Ψc + c0 + c1 (x2 + 1) + c3 (x2 + 1)3 + 6c3 (x2 + 1)x23 ⎦ (5)
0.5σ(x3 − x3 (x2 + 1)2 − x33 )
⎡ 2 2

−αKk p22 + 2α 4 2 2 1 2
η 2 β p22 b1 + 2 c1 ,
1
2 αβ Kk p22 − αη2 Kk β 2 p222 b21 0
−2 2
⎢ 2 2 ⎥
M = ⎣ 1 αβ −2 Kk2 p22 − α2 Kk β 2 p222 b21 , −αKk β −2 p22 + α 2 Kk2 p222 b21 0 ⎦ ≤ 0 (6)
2 η 2η
0 0 0

It can be shown that there exists a γc > 0 such that the ⎡ ⎤


γc 7kΨ
linearized system at the critical point possesses exactly one β −2 (− 5/2 +
3
3/2 )x1
16Ψc 24Ψc
zero eigenvalue at γ = γc with all other eigenvalues in the open C0∗ [x, x, x] = ⎣ c3 x32 + 6c3 x2 x21 ⎦,
left half of the complex plane, and the equilibria are stable for 0.5σ(−x31 − x22 x3 )
γ > γc , but unstable for γ < γc [18]. The subcritical pitchfork
bifurcation solutions are born at the following critical point ⎡ ⎤
2
0
γ = γc = √
c0 +c1 +c3
, C0∗ [r∗ , r∗ , r∗ ] = ⎣ 0 ⎦.
Rc = 0, Φc = 1, −0.5σ
Ψc = c0 + c1 + c3 .
The stabilizing controllers can be synthesized from theo-
Consider the Moore-Greizer model with feedback control rems in [2], [9]. We have the following sufficient conditions
law u = kΨ ∇Ψ = kΨ (Ψ − Ψc ) , then the system (7) can be for a class of stabilizing controllers u = kΨ (x1 − Ψc ) with
rewritten in the following form with the disturbance input w: w = 0 [3], [18].
ẋ = f (x) + g(x)u + b(x)w = f ∗ (x) + bw
(8) Theorem 3.1: The origin of the nonlinear control system
z = cx
(8) is locally asymptotically stable if the following condition
where f ∗ (x) = f (x) + g(x)u, b = [b1 , 0, 0]T with b1 ∈ R, holds:
c = [c1 , 0, 0]T with c1 ∈ R, and  
⎡ ⎤ ⎡ ⎤ 1 3 3 7 3
x1 Ψ − Ψc 0 < Kk < , i.e., − < kΨ < −
6 11 11 66 11
x = ⎣ x2 ⎦ = ⎣ Φ − Φc ⎦ ;
x3 A 

⎡ ⎤ B. Robust design of the axial flow compressor control


−1 √
x1 + Ψe
g(x) = ⎣ 0 ⎦ , ∇γ = γ − γc We shall construct the Lyapunov function for system (8)
β2 with w = 0 in order to calculate the local L2 gain result for
0
the controlled axial flow compressor. Following the Algorithm
and f (x) is given by Equation (5).
(2.3) stated in Sec. 2, the matrix P for system (8) is given by:
Take ∗ = [0, 0, 1], and r∗ = [0, 0, 1]T , and the linear,
⎡ ⎤
quadratic, and cubic terms of its Taylor series are given by β 2 p22 − K2k p22 0
⎡ ⎤ ⎡
1+k Ψ3/2
⎤ P = ⎣ − K2k p22 p22 0 ⎦
− β 2ΨΨcc β −2 0 −β −2 Kk β −2 0
⎢ ⎥ 0 0 1
L∗0 = ⎣ −1 0 0 ⎦=⎣ −1 0 0 ⎦
0 0 0 0 0 0 where p22 = σ2 ( 9K
1
− 23 ).
k
⎡ ⎤ Now we are in the position to prove the following theorem.
0 −1 0
L∗−
0 = ⎣ β 2
−K k 0 ⎦ Theorem 3.2: Let u = kΨ ∇Ψ = kΨ (Ψ − Ψc ) be a class of
0 0 0 stabilizing controllers for the nonlinear control system (8) with
1+kΨ Ψ3/2 −1/2 −3/2 norm-bounded uncertainty signal w, and b(x) = [b1 , 0, 0]T ,
where Kk = Ψc
c
, i.e., kΨ = Kk Ψc − Ψc . c =[c1 , 0, 0]T x, then the best robust controller is kΨ =
⎡ γc Kk 1 ⎤ 7 3
β −2 ( 3/2 −2Ψc + 2
2Ψ2c )x1
− 66 11 = −0.05539.
8Ψc
Q∗0 [x, x] = ⎣ 3c3 (2x23 + x22 ) ⎦, PROOF:
−σx3 x2 Let
⎡ ⎤ V (x) = αxP xT
0
Q∗0 [r∗ , r∗ ] = ⎣ 6c3 ⎦ , with α > 0 be the Lyapunov candidate for the solution of
0 the HJB inequality in Theorem 2.4 or Algorithm 2.5, then

248
kΨ=−0.05539 kΨ=−0.1
0.04 0.04

0.02 0.02

0 0
c

c
−0.02 −0.02
x =Ψ−Ψ

x =Ψ−Ψ
−0.04 −0.04
1

1
−0.06 −0.06

−0.08 −0.08

−0.1 −0.1
0 100 200 300 400 500 0 100 200 300 400 500
Time (second) Time (second)
(a)kΨ = −0.05539 (b)kΨ = −0.10

Fig. 2. Comparison of state x1 between two controllers with small uncertainty w = 0.3 starting from the 0th second and lasting for 1 second.

kΨ=−0.05539 kΨ=−0.1
20 −100

−80

15
x1=Ψ−Ψc −60
x1=Ψ−Ψc
−40
10
−20
x1 and x2

2
x and x

5 0
1

20

0
40
x =Φ−Φ
2 c
60
−5
x2=Φ−Φc 80

−10 100
0 10 20 30 40 50 60 70 80 90
0 100 200 300 400 500
Time (second) Time (second)
(a)kΨ = −0.05539 (b)kΨ = −0.10

Fig. 3. Comparison of states x1 and x2 between two controllers with large uncertainty w = 20 lasting for 10 seconds. For kΨ = −0.14243, only first 90
seconds is shown since the system becomes unstable.

the derivative of the Lyapunov function along the system to the stable subspace of L∗0 :
trajectories (8) is given by
−Kk + j 4β 2 − Kk2
⎡ ⎤ λ1 = ,
 −Kk p22 1 −2 2
Kk p22 0 2β 2
∂V  2β
= xT ⎣ 12 β −2 Kk2 p22 −β −2 Kk p22 0 ⎦ x+O(x3 )  T
∂x f ∗ (x) Kk − j 4β 2 − Kk2
0 0 0 q1 = , 1, 0
2β 2
3
where O(x ) represent the third and higher order terms that,
from Theorem 2.4, do not need to be included in the local L2 −Kk − j 4β 2 − Kk2
λ2 = ,
gain analysis. The matrix M is given by Inequality (6). 2β 2
We can derive the result from Theorem 2.4 with numerical  T
method. Alternatively we can show it by Algorithm 2.5. In this Kk + j 4β 2 − Kk2
q2 = , 1, 0
case, we have the eigenvalues and eigenvectors corresponding 2β 2

249
Then we have derived through searching among a given class of stabilizing
Kk p22 Kk2 bifurcation controls based on the L2 gain value.
L1 = − (2 − ),
β2 β2 ACKNOWLEDGMENT
and This research is supported in part by NSERC Discovery
Kk2 c21
B1 = p222 b21 (β 2 − ), C1 = . Grant and NSERC PGS-D Scholarship.
4 β2
R EFERENCES
Thus η can be estimated by
 [1] E. H. Abed, and J.-H. Fu, “Local feedback stabilization and bifurcation
√ 2 Kk2 control, I. Hopf bifurcation”, Syst. Contr. Lett., 7(1),11-17, 1986.
B1 C1 β 1− 4β 2 β2 [2] E.H. Abed, and J.-H. Fu, “Local feedback stabilization and bifurcation
η̂1 = −2 = Kk2
|b 1 c1 |≈ |b1 c1 | control, II. Stationary bifurcation”, Syst. Contr. Lett., 8(5), 467-473, 1987.
L1 Kk (1 − ) Kk
2β 2 [3] X. Chen, G. Gu, P. Martin, and K. Zhou, “Rotating stall control via
bifurcation stabilization”, Automatica, 34(4), 437-443, 1998.
Since 0 < Kk < 16 , the minimum value of above estimation is
[4] X. Chen, A.G. Sparks, K. Zhou, and G. Gu, “Local Robustness of
1 7 3
Stationary Bifurcation Control”, Latin American Applied Research, 31(3),
achieved when Kk = 6 , i.e., kΨ = − 66 11 = −0.05539. It pp. 163-170, 2001.
is also the best robust controller since it has the least L2 gain. [5] X. Chen, A. Tahmasebi, and G. Gu, “Local Robustness of Bifurcation
Stabilization with Application to Jet Engine Control”, in Bifurcation
Same conclusion can be derived from η̂2 . Control: Theory and Applications, edited by R. Chen et al, Spring-Verlag,
2003.
To demonstrate the effectiveness of the L2 gain method, [6] D. Fontaine, S. Liao, J. Paduano, and P.V. Kokotovic “Nonlinear control
controllers with different L2 gains are compared under differ- expreriments on an axial flow compressor, ”IEEE Trans. Contr. Syst.
Technol., 12(5), 683-693, 2004.
ent uncertainties. Firstly, we compare two controllers: kΨ = [7] J.-H. Fu, and E.H. Abed,“Families of Lyapunov functions for nonlinear
−0.05539 and kΨ = −0.1, which represent the upper bound systems in critical cases”, IEEE Trans. Automat. Contr., 38(1), 3-16, 1993.
and median (approximately) of all stabilizing controllers. We [8] J.T. Gravdahl, and O. Egeland, Compressor Surge and Rotating Stall:
Modeling and Control, Springer, New York, 1999.
assume that x(0) = 0. For simplicity, small uncertainty w is [9] G. Gu, X. Chen, A. Sparks, and S. Banda, “Bifurcation stabilization with
set as a rectangular signal with amplitude 0.3 and duration local output feedback”, SIAM J. Contr. Optim., 37(3), 934-956, 1999.
lasting for 1 second. Transient responses of the state x1 [10] G. Iooss, and D.D. Joseph, Elementary Stability and Bifurcation Theory,
Springer-Verlag, New York, 1980.
of the two controllers are depicted in Fig. (2). It is clear [11] W. Kang, “Bifurcation and normal form of nonlinear control systems:
that oscillation of the controller kΨ = −0.05539 with the Part I and Part II” SIAM J. Contr. Optim., 36, 193-232, 1998.
smaller L2 gain is vanishing very quickly, while another is [12] W. Kang, “Bifurcation control via state feedback for systems with a
single uncontrollable mode,” SIAM J. Contr. Optim., 38, 1428-1452, 2000.
lasting for a considerable long time. This result is comparable [13] T. Kim, and E.H. Abed, “ Stationary bifurcation control of systems with
to that of [18] where the authors show that the controller uncontrollable linearization”, Int. J. Control, 74(5), 445-452, 2001.
kΨ = −0.05539 is capable of tolerating more uncertainties [14] H.K. Khalil, Nonlinear Systems, 3rd Edition, Prentice Hall, New Jersey,
2002.
analyzed by projection method. Secondly, another comparison [15] F.E. McCaughan, “Bifurcation analysis of axial flow compressor stabil-
between two controllers with large uncertainty w = 30 of ity,” SIAM J. Appl. Math.,20,1232-1253, 1990.
duration 5 seconds is also plotted in Fig. (3). The system [16] F.K. Moore, and E.M. Greitzer, “A Theory of Post-stall Transients in
Axial Compression Systems, Parts I & II”, ASME J. of Eng. For Gas
with the controller kΨ = −0.05539 is stable, while it becomes Turbines and Power, 108, 68-97, 1986.
unstable with the controller kΨ = −0.10. [17] A.J. van der Schaft, L2 -gain and Passivity Techniques in Nonlinear
It should be remarked that this L2 gain method could Control, Springer-Verlag: London, 1996.
[18] A. Tahmasebi, and X. Chen, “Robustness of rotating stall control for
give a specific solution to the most robust controller from a axial-flow compressor”, ASME J. Dyn. Syst., Meas., Contr., 125, 424-428,
family of stabilizing controllers, which has advantages over the 2003.
projection method used in [18]. For projection method, it only [19] H.O. Wang, R.A. Adomaitis, and E.H. Abed, “Nonlinear analysis and
control of rotating stall in axial flow compressor,” in Proc. American
gives a range of uncertainty that each controller can tolerate, Contr. Conf. pp. 2317-2321, 1994.
we can not distinguish them from each other unless we resort [20] Y. Wang, and R. Murray, “Feedback stabilization of bifurcations in
to graphic illustrations. multivariable nonlinear systems-Part I: Equilibrium Bifurcations”, Int. J.
of Robust Nonlinear Control, 17, 265-293, 2007.
IV. C ONCLUSION [21] Y. Wang, and R. Murray,“Feedback stabilization of bifurcations in
multivariable nonlinear systems-Part I: Hopf Bifurcations”, Int. J. of
In this paper, robust analysis results based on the L2 gain Robust Nonlinear Control, 17, 295-326, 2007.
approach are derived for a class of stabilizing controllers of [22] M. Xiao, and T. Basar, “Analysis and control of multimode axial flow
compression system models”, ASME J. Dyn. Syst., Meas., Contr., 122,
an axial-flow compressor system with the linearized system 393-401, 2000.
at the equilibrium point possessing a simple zero eigenvalue [23] T. Yang and X. Chen “Local L2 gain of bifurcation stabilization,”
and all other eigenvalues in the open left half of the complex Automatica, to appear, 2008.
plane. In this approach, a class of local feedback controllers is
first synthesized by using the formulae derived from projection
method. A family of Lyapunov functions is then constructed in
the critical case in terms of the Taylor series expansion. Then
the local L2 gain for controlled system has been calculated
through an algorithm. Finally, the most robust controller is

250

You might also like